README.md · Qwen3.6-27B-NVFP4

1

---

2

library_name: transformers

3

license: apache-2.0

4

license_link: https://huggingface.co/Qwen/Qwen3.6-27B/blob/main/LICENSE

5

pipeline_tag: image-text-to-text

6

base_model:

7

- Qwen/Qwen3.6-27B

8

tags:

9

- unsloth

10

- qwen

11

- qwen3_5

12

---

13

# Read our How to [Run Qwen3.6 NVFP4 Guide!](https://unsloth.ai/docs/models/qwen3.6#nvfp4)

14

<div>

15

<p style="margin: 0 0 0px 0; margin-top: 0px;">

16

    <em>See <a href="https://unsloth.ai/docs/basics/unsloth-dynamic-v2.0-gguf">Unsloth Dynamic 2.0 GGUFs</a> for our quantization benchmarks.</em>

17

</p>

18

<div style="display: flex; gap: 5px; align-items: center; margin-bottom: 0px;">

19

<a href="https://github.com/unslothai/unsloth/">

20

<img src="https://github.com/unslothai/unsloth/raw/main/images/unsloth%20new%20logo.png" width="133">

21

</a>

22

<a href="https://discord.gg/unsloth">

23

<img src="https://github.com/unslothai/unsloth/raw/main/images/Discord%20button.png" width="173">

24

</a>

25

<a href="https://unsloth.ai/docs/models/qwen3.6">

26

      <img src="https://raw.githubusercontent.com/unslothai/unsloth/refs/heads/main/images/documentation%20green%20button.png" width="143">

27

</a>

28

</div>

29

30

* **2.5x faster** throughput than other NVFP4 quants.

31

* This is an Unsloth NVFP4 quantized checkpoint calibrated on a mixture of our Unsloth dataset + UltraChat dataset.

32

* Works on a 24GB VRAM GPU. Benchmarks on 1xB200 128 concurrency.

33

34

<img width="800px" src="https://unsloth.ai/docs/~gitbook/image?url=https%3A%2F%2F3215535692-files.gitbook.io%2F%7E%2Ffiles%2Fv0%2Fb%2Fgitbook-x-prod.appspot.com%2Fo%2Fspaces%252FxhOjnexMCB3dmuQFQ2Zq%252Fuploads%252F8zM7cg0Xgo2pAlop1iPW%252F01_qwen36_combined_throughput.png%3Falt%3Dmedia%26token%3D23131a96-4c41-42d7-aa16-5bfb21bb44a3&width=768&dpr=3&quality=100&sign=ee3c70f9&sv=2">

35

36

37

## NVFP4 Accuracy Benchmarks

38

39

For accuracy benchmarks, we conducted MMLU-Pro, AIME 2025, GPQA for FP8, BF16, NVIDIA's NVFP4 and our NVFP4s - we show our faster quants do similarly on all:

40

41

42

| -------- | -------: | ----: | --------: |

43

| Unsloth NVFP4 | 86.25 | 86.34 | 93.12 |

44

| NVIDIA NVFP4 | 85.96 | 86.87 | 93.12 |

45

| FP8 | 86.11 | 86.87 | 93.75 |

46

| BF16 | 85.96 | 88.13 | 93.33 |

47

48

Read all benchmarks in our [NVFP4 blog](https://unsloth.ai/docs/models/qwen3.6#nvfp4)

49

50

## vLLM Run Instructions

51

52

To install vLLM in a separate venv:

53

```shell

54

uv venv unsloth-nvfp4-env --python 3.13

55

source unsloth-nvfp4-env/bin/activate

56

uv pip install "vllm>=0.25.0" "flashinfer-python>=0.6.13" "nvidia-cutlass-dsl>=4.5.2" \

57

--torch-backend=auto

58

```

59

Then to serve the 27B NVFP4 quant:

60

```

61

vllm serve unsloth/Qwen3.6-27B-NVFP4 \

62

--speculative-config '{"method": "mtp", "num_speculative_tokens": 2}'

63

```

64

65

Also do **NOT** use the Marlin backend since it's 2x slower - use the native vLLM or cute-DSL / CUTLASS / flashinfer_trtllm backends!

66

67

|---|---|---|---|

68

| nvidia 27B | marlin (auto) | 115.6 | 2,403 |

69

| unsloth 27B | cute-DSL (auto) | 125.9 | **6,863** |

70

| nvidia 35B-A3B | marlin (auto) | 240.8 | 8,721 |

71

| unsloth 35B-A3B | cute-DSL + trtllm (auto) | 295.2 | **15,636** |

72

73

## DGX Spark

74

75

You must use the below or you will get **2x slower** inference!

76

```shell

77

export CUTE_DSL_ARCH=sm_121a

78

vllm serve unsloth/Qwen3.6-27B-NVFP4 --moe-backend flashinfer_b12x

79

```

80

81

## Multi-Token Prediction (MTP)

82

83

This checkpoint includes the MTP module, so it can act as its own speculative draft for faster decoding, but somewhat less throughput

84

85

```shell

86

vllm serve unsloth/Qwen3.6-27B-NVFP4 \

87

--speculative-config '{"method": "mtp", "num_speculative_tokens": 2}'

88

```

89

90

[![Qwen Chat](https://img.shields.io/badge/%F0%9F%92%9C%EF%B8%8F%20Qwen%20Chat%20-536af5)](https://chat.qwen.ai)

91

92

> [!Note]

93

> This repository contains model weights and configuration files for the post-trained model in the Hugging Face Transformers format.

94

>

95

> These artifacts are compatible with Hugging Face Transformers, vLLM, SGLang, KTransformers, etc.

96

97

Following the February release of the Qwen3.5 series, we're pleased to share the first open-weight variant of Qwen3.6. Built on direct feedback from the community, Qwen3.6 prioritizes stability and real-world utility, offering developers a more intuitive, responsive, and genuinely productive coding experience.

98

99

## Qwen3.6 Highlights

100

101

This release delivers substantial upgrades, particularly in

102

103

- **Agentic Coding:** the model now handles frontend workflows and repository-level reasoning with greater fluency and precision.

104

- **Thinking Preservation:** we've introduced a new option to retain reasoning context from historical messages, streamlining iterative development and reducing overhead.

105

106

![Benchmark Results](https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3.6/Figures/qwen3.6_27b_score.png)

107

108

For more details, please refer to our blog post [Qwen3.6-27B](https://qwen.ai/blog?id=qwen3.6-27b).

109

110

## Model Overview

111

112

- Type: Causal Language Model with Vision Encoder

113

- Training Stage: Pre-training & Post-training

114

- Language Model

115

- Number of Parameters: 27B

116

- Hidden Dimension: 5120

117

- Token Embedding: 248320 (Padded)

118

- Number of Layers: 64

119

- Hidden Layout: 16 × (3 × (Gated DeltaNet → FFN) → 1 × (Gated Attention → FFN))

120

- Gated DeltaNet:

121

- Number of Linear Attention Heads: 48 for V and 16 for QK

122

- Head Dimension: 128

123

- Gated Attention:

124

- Number of Attention Heads: 24 for Q and 4 for KV

125

- Head Dimension: 256

126

- Rotary Position Embedding Dimension: 64

127

- Feed Forward Network:

128

- Intermediate Dimension: 17408

129

- LM Output: 248320 (Padded)

130

- MTP: trained with multi-steps

131

- Context Length: 262,144 natively and extensible up to 1,010,000 tokens.

132

133

134

## Original Qwen3.6 BF16 Reference Benchmarks

135

136

### Language

137

138

<div style="font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,sans-serif;max-width:1000px;margin:0 auto;padding:16px 0">

139

<table style="width:100%;border-collapse:collapse;font-size:13px">

140

<thead><tr>

141

<th style="padding:10px 7px;text-align:left;font-weight:600;border-bottom:2px solid #7c3aed;color:#7c3aed"></th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.5-27B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.5-397B-A17B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Gemma4-31B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Claude 4.5 Opus</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.6-35B-A3B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.6-27B</th></tr></thead>

142

<tbody>

143

<tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#7c3aed;border-bottom:1px solid rgba(124, 58, 237, 0.2);background:rgba(124, 58, 237, 0.1)">Coding Agent</td></tr>

144

<tr>

145

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SWE-bench Verified</td>

146

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">75.0</td>

147

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">76.2</td>

148

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">52.0</td>

149

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.9</td>

150

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">73.4</td>

151

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">77.2</td>

152

</tr>

153

<tr>

154

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SWE-bench Pro</td>

155

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">51.2</td>

156

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">50.9</td>

157

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">35.7</td>

158

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">57.1</td>

159

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">49.5</td>

160

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">53.5</td>

161

</tr>

162

<tr>

163

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SWE-bench Multilingual</td>

164

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">69.3</td>

165

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">69.3</td>

166

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">51.7</td>

167

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">77.5</td>

168

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">67.2</td>

169

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">71.3</td>

170

</tr>

171

<tr>

172

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">Terminal-Bench 2.0</td>

173

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">41.6</td>

174

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">52.5</td>

175

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">42.9</td>

176

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">59.3</td>

177

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">51.5</td>

178

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">59.3</td>

179

</tr>

180

<tr>

181

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SkillsBench <sub><small>Avg5</small></sub></td>

182

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">27.2</td>

183

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">30.0</td>

184

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">23.6</td>

185

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">45.3</td>

186

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">28.7</td>

187

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">48.2</td>

188

</tr>

189

<tr>

190

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">QwenWebBench</td>

191

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">1068</td>

192

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">1186</td>

193

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">1197</td>

194

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">1536</td>

195

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">1397</td>

196

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">1487</td>

197

</tr>

198

<tr>

199

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">NL2Repo</td>

200

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">27.3</td>

201

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">32.2</td>

202

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">15.5</td>

203

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">43.2</td>

204

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">29.4</td>

205

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">36.2</td>

206

</tr>

207

<tr>

208

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">Claw-Eval <sub><small>Avg</small></sub></td>

209

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">64.3</td>

210

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">70.7</td>

211

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">48.5</td>

212

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">76.6</td>

213

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">68.7</td>

214

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">72.4</td>

215

</tr>

216

<tr>

217

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">Claw-Eval <sub><small>Pass^3</small></sub></td>

218

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">46.2</td>

219

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">48.1</td>

220

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">25.0</td>

221

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">59.6</td>

222

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">50.0</td>

223

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">60.6</td>

224

</tr>

225

<tr>

226

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">QwenClawBench</td>

227

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">52.2</td>

228

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">51.8</td>

229

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">41.7</td>

230

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">52.3</td>

231

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">52.6</td>

232

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">53.4</td>

233

</tr>

234

<tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#7c3aed;border-bottom:1px solid rgba(124, 58, 237, 0.2);background:rgba(124, 58, 237, 0.1)">Knowledge</td></tr>

235

<tr>

236

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMLU-Pro</td>

237

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.1</td>

238

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.8</td>

239

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">85.2</td>

240

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">89.5</td>

241

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">85.2</td>

242

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.2</td>

243

</tr>

244

<tr>

245

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMLU-Redux</td>

246

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">93.2</td>

247

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">94.9</td>

248

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">93.7</td>

249

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">95.6</td>

250

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">93.3</td>

251

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">93.5</td>

252

</tr>

253

<tr>

254

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SuperGPQA</td>

255

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">65.6</td>

256

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">70.4</td>

257

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">65.7</td>

258

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">70.6</td>

259

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">64.7</td>

260

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">66.0</td>

261

</tr>

262

<tr>

263

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">C-Eval</td>

264

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">90.5</td>

265

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">93.0</td>

266

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">82.6</td>

267

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.2</td>

268

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">90.0</td>

269

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">91.4</td>

270

</tr>

271

<tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#7c3aed;border-bottom:1px solid rgba(124, 58, 237, 0.2);background:rgba(124, 58, 237, 0.1)">STEM & Reasoning</td></tr>

272

<tr>

273

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">GPQA Diamond</td>

274

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">85.5</td>

275

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">88.4</td>

276

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.3</td>

277

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.0</td>

278

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.0</td>

279

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.8</td>

280

</tr>

281

<tr>

282

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">HLE</td>

283

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">24.3</td>

284

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">28.7</td>

285

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">19.5</td>

286

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">30.8</td>

287

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">21.4</td>

288

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">24.0</td>

289

</tr>

290

<tr>

291

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">LiveCodeBench v6</td>

292

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.7</td>

293

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">83.6</td>

294

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.0</td>

295

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.8</td>

296

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.4</td>

297

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">83.9</td>

298

</tr>

299

<tr>

300

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">HMMT Feb 25</td>

301

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.0</td>

302

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">94.8</td>

303

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">88.7</td>

304

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.9</td>

305

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">90.7</td>

306

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">93.8</td>

307

</tr>

308

<tr>

309

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">HMMT Nov 25</td>

310

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">89.8</td>

311

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.7</td>

312

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.5</td>

313

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">93.3</td>

314

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">89.1</td>

315

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">90.7</td>

316

</tr>

317

<tr>

318

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">HMMT Feb 26</td>

319

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.3</td>

320

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.9</td>

321

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">77.2</td>

322

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">85.3</td>

323

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">83.6</td>

324

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.3</td>

325

</tr>

326

<tr>

327

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">IMOAnswerBench</td>

328

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">79.9</td>

329

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.9</td>

330

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">74.5</td>

331

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.0</td>

332

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">78.9</td>

333

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.8</td>

334

</tr>

335

<tr>

336

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">AIME26</td>

337

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.6</td>

338

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">93.3</td>

339

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">89.2</td>

340

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">95.1</td>

341

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.7</td>

342

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">94.1</td>

343

</tr>

344

</tbody>

345

</table>

346

347

<p style="margin-top:12px;font-size:10px;opacity:0.7">

348

* SWE-Bench Series: Internal agent scaffold (bash + file-edit tools); temp=1.0, top_p=0.95, 200K context window. We correct some problematic tasks in the public set of SWE-bench Pro and evaluate all baselines on the refined benchmark.<br/>

349

* Terminal-Bench 2.0: Harbor/Terminus-2 harness; 3h timeout, 32 CPU/48 GB RAM; temp=1.0, top_p=0.95, top_k=20, max_tokens=80K, 256K ctx; avg of 5 runs.<br/>

350

* SkillsBench: Evaluated via OpenCode on 78 tasks (self-contained subset, excluding API-dependent tasks); avg of 5 runs.<br/>

351

* NL2Repo: Others are evaluated via Claude Code (temp=1.0, top_p=0.95, max_turns=900).<br/>

352

* QwenClawBench: A real-user-distribution Claw agent benchmark; temp=0.6, 256K ctx.<br/>

353

* QwenWebBench: An internal front-end code generation benchmark; bilingual (EN/CN), 7 categories (Web Design, Web Apps, Games, SVG, Data Visualization, Animation, and 3D); auto-render + multimodal judge (code/visual correctness); BT/Elo rating system.<br/>

354

* AIME 26: We use the full AIME 2026 (I & II), where the scores may differ from Qwen 3.5 notes.

355

</p>

356

357

</div>

358

359

360

### Vision Language

361

362

<div style="font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,sans-serif;max-width:1000px;margin:0 auto;padding:16px 0">

363

<table style="width:100%;border-collapse:collapse;font-size:13px">

364

<thead><tr>

365

<th style="padding:10px 7px;text-align:left;font-weight:600;border-bottom:2px solid #7c3aed;color:#7c3aed"></th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.5-27B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.5-397B-A17B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Gemma4-31B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Claude 4.5 Opus</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.6-35B-A3B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.6-27B</th></tr></thead>

366

<tbody>

367

<tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#7c3aed;border-bottom:1px solid rgba(124, 58, 237, 0.2);background:rgba(124, 58, 237, 0.1)">STEM & Puzzle</td></tr>

368

<tr>

369

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMMU</td>

370

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">82.3</td>

371

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">85.0</td>

372

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.4</td>

373

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.7</td>

374

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">81.7</td>

375

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">82.9</td>

376

</tr>

377

<tr>

378

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMMU-Pro</td>

379

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">75.0</td>

380

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">79.0</td>

381

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">76.9</td>

382

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">70.6</td>

383

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">75.3</td>

384

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">75.8</td>

385

</tr>

386

<tr>

387

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MathVista <sub><small>mini</small></sub></td>

388

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.8</td>

389

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

390

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">79.3</td>

391

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

392

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.4</td>

393

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.4</td>

394

</tr>

395

<tr>

396

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">DynaMath</td>

397

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.7</td>

398

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.3</td>

399

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">79.5</td>

400

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">79.7</td>

401

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">82.8</td>

402

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">85.6</td>

403

</tr>

404

<tr>

405

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">VlmsAreBlind</td>

406

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">96.9</td>

407

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

408

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.2</td>

409

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

410

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">96.6</td>

411

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">97.0</td>

412

</tr>

413

<tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#7c3aed;border-bottom:1px solid rgba(124, 58, 237, 0.2);background:rgba(124, 58, 237, 0.1)">General VQA</td></tr>

414

<tr>

415

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">RealWorldQA</td>

416

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">83.7</td>

417

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">83.9</td>

418

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">72.3</td>

419

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">77.0</td>

420

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">85.3</td>

421

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.1</td>

422

</tr>

423

<tr>

424

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMStar</td>

425

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">81.0</td>

426

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">83.8</td>

427

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">77.3</td>

428

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">73.2</td>

429

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.7</td>

430

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">81.4</td>

431

</tr>

432

<tr>

433

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMBench<sub><small>EN-DEV-v1.1</small></sub></td>

434

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.6</td>

435

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

436

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">90.9</td>

437

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

438

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.8</td>

439

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.3</td>

440

</tr>

441

<tr>

442

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SimpleVQA</td>

443

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">56.0</td>

444

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">67.1</td>

445

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">52.9</td>

446

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">65.7</td>

447

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">58.9</td>

448

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">56.1</td>

449

</tr>

450

<tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#7c3aed;border-bottom:1px solid rgba(124, 58, 237, 0.2);background:rgba(124, 58, 237, 0.1)">Document Understanding</td></tr>

451

<tr>

452

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">CharXiv <sub><small>RQ</small></sub></td>

453

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">79.5</td>

454

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.8</td>

455

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">67.9</td>

456

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">68.5</td>

457

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">78.0</td>

458

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">78.4</td>

459

</tr>

460

<tr>

461

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">CC-OCR</td>

462

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">81.0</td>

463

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">82.0</td>

464

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">75.7</td>

465

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">76.9</td>

466

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">81.9</td>

467

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">81.2</td>

468

</tr>

469

<tr>

470

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">OCRBench</td>

471

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">89.4</td>

472

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

473

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.1</td>

474

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

475

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">90.0</td>

476

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">89.4</td>

477

</tr>

478

<tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#7c3aed;border-bottom:1px solid rgba(124, 58, 237, 0.2);background:rgba(124, 58, 237, 0.1)">Spatial Intelligence</td></tr>

479

<tr>

480

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">ERQA</td>

481

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">60.5</td>

482

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">67.5</td>

483

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">57.5</td>

484

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">46.8</td>

485

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">61.8</td>

486

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">62.5</td>

487

</tr>

488

<tr>

489

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">CountBench</td>

490

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">97.8</td>

491

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">97.2</td>

492

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">96.1</td>

493

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">90.6</td>

494

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">96.1</td>

495

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">97.8</td>

496

</tr>

497

<tr>

498

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">RefCOCO <sub><small>avg</small></sub></td>

499

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">90.9</td>

500

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.3</td>

501

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

502

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

503

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.0</td>

504

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">92.5</td>

505

</tr>

506

<tr>

507

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">EmbSpatialBench</td>

508

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.5</td>

509

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

510

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

511

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

512

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.3</td>

513

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.6</td>

514

</tr>

515

<tr>

516

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">RefSpatialBench</td>

517

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">67.7</td>

518

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

519

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">4.7</td>

520

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

521

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">64.3</td>

522

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">70.0</td>

523

</tr>

524

<tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#7c3aed;border-bottom:1px solid rgba(124, 58, 237, 0.2);background:rgba(124, 58, 237, 0.1)">Video Understanding</td></tr>

525

<tr>

526

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">VideoMME<sub><small>(w sub.)</sub></small></td>

527

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.0</td>

528

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.5</td>

529

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

530

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">77.7</td>

531

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.6</td>

532

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.7</td>

533

</tr>

534

<tr>

535

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">VideoMMMU</td>

536

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">82.3</td>

537

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.7</td>

538

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">81.6</td>

539

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.4</td>

540

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">83.7</td>

541

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.4</td>

542

</tr>

543

<tr>

544

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MLVU</td>

545

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">85.9</td>

546

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.7</td>

547

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

548

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">81.7</td>

549

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.2</td>

550

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.6</td>

551

</tr>

552

<tr>

553

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MVBench</td>

554

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">74.6</td>

555

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">77.6</td>

556

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

557

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">67.2</td>

558

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">74.6</td>

559

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">75.5</td>

560

</tr>

561

<tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#7c3aed;border-bottom:1px solid rgba(124, 58, 237, 0.2);background:rgba(124, 58, 237, 0.1)">Visual Agent</td></tr>

562

<tr>

563

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">V*</td>

564

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">93.7</td>

565

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">95.8</td>

566

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

567

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">67.0</td>

568

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">90.1</td>

569

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">94.7</td>

570

</tr>

571

<tr>

572

<td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">AndroidWorld</td>

573

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">64.2</td>

574

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

575

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

576

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

577

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>

578

<td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">70.3</td>

579

</tr>

580

</tbody>

581

</table>

582

583

<p style="margin-top:12px;font-size:10px;opacity:0.7">

584

* Empty cells (--) indicate scores not yet available or not applicable.

585

</p>

586

587

</div>

588

589

590

## Quickstart

591

592

For streamlined integration, we recommend using Qwen3.6 via APIs. Below is a guide to use Qwen3.6 via OpenAI-compatible API.

593

594

### Serving Qwen3.6

595

596

Qwen3.6 can be served via APIs with popular inference frameworks.

597

In the following, we show example commands to launch OpenAI-Compatible API servers for Qwen3.6 models.

598

599

> [!Important]

600

> Inference efficiency and throughput vary significantly across frameworks.

601

> We recommend using the latest framework versions to ensure optimal performance and compatibility.

602

> For production workloads or high-throughput scenarios, dedicated serving engines such as SGLang, KTransformers or vLLM are strongly recommended.

603

604

> [!Important]

605

> The model has a default context length of 262,144 tokens.

606

> If you encounter out-of-memory (OOM) errors, consider reducing the context window.

607

> However, because Qwen3.6 leverages extended context for complex tasks, we advise maintaining a context length of at least 128K tokens to preserve thinking capabilities.

608

609

#### SGLang

610

611

[SGLang](https://github.com/sgl-project/sglang) is a fast serving framework for large language models and vision language models.

612

`sglang>=0.5.10` is recommended for Qwen3.6, which can be installed using the following command in a fresh environment:

613

```shell

614

uv pip install sglang[all]

615

```

616

See [its documentation](https://docs.sglang.ai/get_started/install.html) for more details.

617

618

The following will create API endpoints at `http://localhost:8000/v1`:

619

620

- **Standard Version**: The following command can be used to create an API endpoint with maximum context length 262,144 tokens using tensor parallel on 8 GPUs.

621

622

```shell

623

    python -m sglang.launch_server --model-path Qwen/Qwen3.6-27B --port 8000 --tp-size 8 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3

624

```

625

626

- **Tool Use**: To support tool use, you can use the following command.

627

628

```shell

629

    python -m sglang.launch_server --model-path Qwen/Qwen3.6-27B --port 8000 --tp-size 8 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3 --tool-call-parser qwen3_coder

630

```

631

632

- **Multi-Token Prediction (MTP)**: The following command is recommended for MTP:

633

634

```shell

635

    python -m sglang.launch_server --model-path Qwen/Qwen3.6-27B --port 8000 --tp-size 8 --mem-fraction-static 0.8 --context-length 262144 --reasoning-parser qwen3 --speculative-algo NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4

636

```

637

638

For detailed deployment guide, see the [SGLang Qwen3.5 Cookbook](https://lmsysorg.mintlify.app/cookbook/llm/Qwen/Qwen3.5).

639

640

#### vLLM

641

642

[vLLM](https://github.com/vllm-project/vllm) is the recommended serving path for this NVFP4 checkpoint. `vllm>=0.19.0` is recommended for Qwen3.6 NVFP4 artifacts:

643

```shell

644

uv pip install vllm --torch-backend=auto

645

```

646

647

The following command creates an OpenAI-compatible API endpoint at `http://localhost:8000/v1`:

648

649

```shell

650

vllm serve unsloth/Qwen3.6-27B-NVFP4 --trust-remote-code --dtype bfloat16 --max-model-len 4096

651

```

652

653

Increase `--max-model-len` only after checking available GPU memory. For long-context serving, follow the YaRN notes below and keep the model ID pointed at `unsloth/Qwen3.6-27B-NVFP4`.

654

655

#### KTransformers

656

657

[KTransformers](https://github.com/kvcache-ai/ktransformers) is a flexible framework for experiencing cutting-edge LLM inference optimizations with CPU-GPU heterogeneous computing.

658

For running Qwen3.6 with KTransformers, see the [KTransformers Deployment Guide](https://github.com/kvcache-ai/ktransformers/blob/main/doc/en/Qwen3.5.md).

659

660

#### Hugging Face Transformers

661

662

Hugging Face Transformers contains a _lightweight_ server which can be used for quick testing and moderate load deployment.

663

The latest `transformers` is required for Qwen3.6:

664

```shell

665

pip install "transformers[serving]"

666

```

667

See [its documentation](https://huggingface.co/docs/transformers/main/serving) for more details. Please also make sure torchvision and pillow are installed.

668

669

Then, run `transformers serve` to launch a server with API endpoints at `http://localhost:8000/v1`; it will place the model on accelerators if available:

670

```shell

671

transformers serve Qwen/Qwen3.6-27B --port 8000 --continuous-batching

672

```

673

674

### Using Qwen3.6 via the Chat Completions API

675

676

The chat completions API is accessible via standard HTTP requests or OpenAI SDKs.

677

Here, we show examples using the OpenAI Python SDK.

678

679

Before starting, make sure it is installed and the API key and the API base URL is configured, e.g.:

680

```shell

681

pip install -U openai

682

683

# Set the following accordingly

684

export OPENAI_BASE_URL="http://localhost:8000/v1"

685

export OPENAI_API_KEY="EMPTY"

686

```

687

688

> [!Tip]

689

> We recommend using the following set of sampling parameters for generation

690

> - Thinking mode for general tasks: `temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0`

691

> - Thinking mode for precise coding tasks (e.g. WebDev): `temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0`

692

> - Instruct (or non-thinking) mode: `temperature=0.7, top_p=0.80, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0`

693

>

694

> Please note that the support for sampling parameters varies according to inference frameworks.

695

696

> [!Important]

697

> Qwen3.6 models operate in thinking mode by default, generating thinking content signified by `<think>\n...</think>\n\n` before producing the final responses.

698

> To disable thinking content and obtain direct response, refer to the examples [here](#instruct-or-non-thinking-mode).

699

700

701

#### Text-Only Input

702

703

```python

704

from openai import OpenAI

705

# Configured by environment variables

706

client = OpenAI()

707

708

messages = [

709

{"role": "user", "content": "Type \"I love Qwen3.6\" backwards"},

710

]

711

712

chat_response = client.chat.completions.create(

713

model="Qwen/Qwen3.6-27B",

714

messages=messages,

715

max_tokens=81920,

716

temperature=1.0,

717

top_p=0.95,

718

presence_penalty=0.0,

719

extra_body={

720

"top_k": 20,

721

},

722

)

723

print("Chat response:", chat_response)

724

```

725

726

727

#### Image Input

728

729

```python

730

from openai import OpenAI

731

# Configured by environment variables

732

client = OpenAI()

733

734

messages = [

735

{

736

"role": "user",

737

"content": [

738

{

739

"type": "image_url",

740

"image_url": {

741

"url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/CI_Demo/mathv-1327.jpg"

742

}

743

},

744

{

745

"type": "text",

746

                "text": "The centres of the four illustrated circles are in the corners of the square. The two big circles touch each other and also the two little circles. With which factor do you have to multiply the radii of the little circles to obtain the radius of the big circles?\nChoices:\n(A) $\\frac{2}{9}$\n(B) $\\sqrt{5}$\n(C) $0.8 \\cdot \\pi$\n(D) 2.5\n(E) $1+\\sqrt{2}$"

747

}

748

]

749

}

750

]

751

752

response = client.chat.completions.create(

753

model="Qwen/Qwen3.6-27B",

754

messages=messages,

755

max_tokens=81920,

756

temperature=1.0,

757

top_p=0.95,

758

presence_penalty=0.0,

759

extra_body={

760

"top_k": 20,

761

},

762

)

763

print("Chat response:", chat_response)

764

```

765

766

#### Video Input

767

768

```python

769

from openai import OpenAI

770

# Configured by environment variables

771

client = OpenAI()

772

773

messages = [

774

{

775

"role": "user",

776

"content": [

777

{

778

"type": "video_url",

779

"video_url": {

780

"url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/video/N1cdUjctpG8.mp4"

781

}

782

},

783

{

784

"type": "text",

785

                "text": "How many porcelain jars were discovered in the niches located in the primary chamber of the tomb?"

786

}

787

]

788

}

789

]

790

791

# When vLLM is launched with `--media-io-kwargs '{"video": {"num_frames": -1}}'`,

792

# video frame sampling can be configured via `extra_body` (e.g., by setting `fps`).

793

# This feature is currently supported only in vLLM.

794

#

795

# By default, `fps=2` and `do_sample_frames=True`.

796

# With `do_sample_frames=True`, you can customize the `fps` value to set your desired video sampling rate.

797

response = client.chat.completions.create(

798

model="Qwen/Qwen3.6-27B",

799

messages=messages,

800

max_tokens=81920,

801

temperature=1.0,

802

top_p=0.95,

803

presence_penalty=0.0,

804

extra_body={

805

"top_k": 20,

806

"mm_processor_kwargs": {"fps": 2, "do_sample_frames": True},

807

},

808

)

809

810

print("Chat response:", chat_response)

811

```

812

813

814

#### Instruct (or Non-Thinking) Mode

815

816

> [!Important]

817

> Qwen3.6 does not officially support the soft switch of Qwen3, i.e., `/think` and `/nothink`.

818

819

Qwen3.6 will think by default before response.

820

You can obtain direct response from the model without thinking by configuring the API parameters.

821

For example,

822

```python

823

from openai import OpenAI

824

# Configured by environment variables

825

client = OpenAI()

826

827

messages = [

828

{

829

"role": "user",

830

"content": [

831

{

832

"type": "image_url",

833

"image_url": {

834

"url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.6/demo/RealWorld/RealWorld-04.png"

835

}

836

},

837

{

838

"type": "text",

839

"text": "Where is this?"

840

}

841

]

842

}

843

]

844

845

chat_response = client.chat.completions.create(

846

model="Qwen/Qwen3.6-27B",

847

messages=messages,

848

max_tokens=32768,

849

temperature=0.7,

850

top_p=0.8,

851

presence_penalty=1.5,

852

extra_body={

853

"top_k": 20,

854

"chat_template_kwargs": {"enable_thinking": False},

855

},

856

)

857

print("Chat response:", chat_response)

858

```

859

860

> [!Note]

861

> If you are using APIs from Alibaba Cloud Model Studio, in addition to changing `model`, please use `"enable_thinking": False` instead of `"chat_template_kwargs": {"enable_thinking": False}`.

862

863

#### Preserve Thinking

864

865

By default, only the thinking blocks generated in handling the latest user message is retained, resulting in a pattern commonly as interleaved thinking.

866

Qwen3.6 has been additionally trained to preserve and leverage thinking traces from historical messages.

867

You can enable this behavior by setting the `preserve_thinking` option:

868

```python

869

from openai import OpenAI

870

# Configured by environment variables

871

client = OpenAI()

872

873

messages = [...]

874

875

chat_response = client.chat.completions.create(

876

model="Qwen/Qwen3.6-27B",

877

messages=messages,

878

max_tokens=32768,

879

temperature=0.6,

880

top_p=0.95,

881

presence_penalty=0.0,

882

extra_body={

883

"top_k": 20,

884

"chat_template_kwargs": {"preserve_thinking": True},

885

},

886

)

887

print("Chat response:", chat_response)

888

```

889

890

> [!Note]

891

> If you are using APIs from Alibaba Cloud Model Studio, in addition to changing `model`, please use `"preserve_thinking": True` instead of `"chat_template_kwargs": {"preserve_thinking": False}`.

892

893

894

This capability is particularly beneficial for agent scenarios, where maintaining full reasoning context can enhance decision consistency and, in many cases, reduce overall token consumption by minimizing redundant reasoning. Additionally, it can improve KV cache utilization, optimizing inference efficiency in both thinking and non-thinking modes.

895

896

897

## Agentic Usage

898

899

Qwen3.6 excels in tool calling capabilities.

900

901

### Qwen-Agent

902

903

We recommend using [Qwen-Agent](https://github.com/QwenLM/Qwen-Agent) to quickly build Agent applications with Qwen3.6.

904

905

To define the available tools, you can use the MCP configuration file, use the integrated tool of Qwen-Agent, or integrate other tools by yourself.

906

```python

907

import os

908

from qwen_agent.agents import Assistant

909

910

# Define LLM

911

# Using Alibaba Cloud Model Studio

912

llm_cfg = {

913

# Use the OpenAI-compatible model service provided by DashScope:

914

'model': 'qwen3.6-27b',

915

'model_type': 'qwenvl_oai',

916

'model_server': 'https://dashscope.aliyuncs.com/compatible-mode/v1',

917

'api_key': os.getenv('DASHSCOPE_API_KEY'),

918

919

'generate_cfg': {

920

'use_raw_api': True,

921

# When using Dash Scope OAI API, pass the parameter of whether to enable thinking mode in this way

922

'extra_body': {

923

'enable_thinking': True,

924

'preserve_thinking': True,

925

},

926

},

927

}

928

929

# Using OpenAI-compatible API endpoint.

930

# functionality of the deployment frameworks and let Qwen-Agent automate the related operations.

931

#

932

# llm_cfg = {

933

# # Use your own model service compatible with OpenAI API by vLLM/SGLang:

934

# 'model': 'unsloth/Qwen3.6-27B-NVFP4',

935

# 'model_type': 'qwenvl_oai',

936

# 'model_server': 'http://localhost:8000/v1', # api_base

937

# 'api_key': 'EMPTY',

938

#

939

# 'generate_cfg': {

940

# 'use_raw_api': True,

941

# # When using vLLM/SGLang OAI API, pass the parameter of whether to enable thinking mode in this way

942

# 'extra_body': {

943

# 'chat_template_kwargs': {'enable_thinking': True, 'preserve_thinking': True}

944

# },

945

# },

946

# }

947

948

# Define Tools

949

tools = [

950

{'mcpServers': { # You can specify the MCP configuration file

951

"filesystem": {

952

"command": "npx",

953

"args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/xxxx/Desktop"]

954

}

955

}

956

}

957

]

958

959

# Define Agent

960

bot = Assistant(llm=llm_cfg, function_list=tools)

961

962

# Streaming generation

963

messages = [{'role': 'user', 'content': 'Help me organize my desktop.'}]

964

for responses in bot.run(messages=messages):

965

pass

966

print(responses)

967

968

# Streaming generation

969

messages = [{'role': 'user', 'content': 'Develop a dog website and save it on the desktop'}]

970

for responses in bot.run(messages=messages):

971

pass

972

print(responses)

973

```

974

975

### Qwen Code

976

977

978

[Qwen Code](https://github.com/QwenLM/qwen-code) is an open-source AI agent for the terminal, optimized for Qwen models. It helps you understand large codebases, automate tedious work, and ship faster.

979

980

For more information, please refer to [Qwen Code](https://qwenlm.github.io/qwen-code-docs/).

981

982

## Processing Ultra-Long Texts

983

984

Qwen3.6 natively supports context lengths of up to 262,144 tokens.

985

For long-horizon tasks where the total length (including both input and output) exceeds this limit, we recommend using RoPE scaling techniques to handle long texts effectively., e.g., YaRN.

986

987

YaRN is currently supported by several inference frameworks, e.g., `transformers`, `vllm`, `ktransformers` and `sglang`.

988

In general, there are two approaches to enabling YaRN for supported frameworks:

989

990

- Modifying the model configuration file:

991

In the `config.json` file, change the `rope_parameters` fields in `text_config` to:

992

```json

993

{

994

"mrope_interleaved": true,

995

"mrope_section": [

996

11,

997

11,

998

10

999

],

1000

"rope_type": "yarn",

1001

"rope_theta": 10000000,

1002

"partial_rotary_factor": 0.25,

1003

"factor": 4.0,

1004

"original_max_position_embeddings": 262144,

1005

}

1006

```

1007

1008

- Passing command line arguments:

1009

1010

For `vllm`, you can use

1011

```shell

1012

    VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 vllm serve ... --hf-overrides '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}' --max-model-len 1010000

1013

```

1014

1015

For `sglang` and `ktransformers`, you can use

1016

```shell

1017

    SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 python -m sglang.launch_server ... --json-model-override-args '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}' --context-length 1010000

1018

```

1019

1020

> [!NOTE]

1021

> All the notable open-source frameworks implement static YaRN, which means the scaling factor remains constant regardless of input length, **potentially impacting performance on shorter texts.**

1022

> We advise modifying the `rope_parameters` configuration only when processing long contexts is required.

1023

> It is also recommended to modify the `factor` as needed. For example, if the typical context length for your application is 524,288 tokens, it would be better to set `factor` as 2.0.

1024

1025

## Best Practices

1026

1027

To achieve optimal performance, we recommend the following settings:

1028

1029

1. **Sampling Parameters**:

1030

- We suggest using the following sets of sampling parameters depending on the mode and task type:

1031

- **Thinking mode for general tasks**:

1032

`temperature=1.0`, `top_p=0.95`, `top_k=20`, `min_p=0.0`, `presence_penalty=0.0`, `repetition_penalty=1.0`

1033

- **Thinking mode for precise coding tasks (e.g., WebDev)**:

1034

`temperature=0.6`, `top_p=0.95`, `top_k=20`, `min_p=0.0`, `presence_penalty=0.0`, `repetition_penalty=1.0`

1035

- **Instruct (or non-thinking) mode**:

1036

`temperature=0.7`, `top_p=0.80`, `top_k=20`, `min_p=0.0`, `presence_penalty=1.5`, `repetition_penalty=1.0`

1037

   - For supported frameworks, you can adjust the `presence_penalty` parameter between 0 and 2 to reduce endless repetitions. However, using a higher value may occasionally result in language mixing and a slight decrease in model performance.

1038

1039

2. **Adequate Output Length**: We recommend using an output length of 32,768 tokens for most queries. For benchmarking on highly complex problems, such as those found in math and programming competitions, we suggest setting the max output length to 81,920 tokens. This provides the model with sufficient space to generate detailed and comprehensive responses, thereby enhancing its overall performance.

1040

1041

3. **Standardize Output Format**: We recommend using prompts to standardize model outputs when benchmarking.

1042

- **Math Problems**: Include "Please reason step by step, and put your final answer within \boxed{}." in the prompt.

1043

   - **Multiple-Choice Questions**: Add the following JSON structure to the prompt to standardize responses: "Please show your choice in the `answer` field with only the choice letter, e.g., `"answer": "C"`."

1044

1045

4. **Long Video Understanding**: To optimize inference efficiency for plain text and images, the `size` parameter in the released `video_preprocessor_config.json` is conservatively configured. It is recommended to set the `longest_edge` parameter in the video_preprocessor_config file to 469,762,048 (corresponding to 224k video tokens) to enable higher frame-rate sampling for hour-scale videos and thereby achieve superior performance. For example,

1046

```json

1047

{"longest_edge": 469762048, "shortest_edge": 4096}

1048

```

1049

1050

    Alternatively, override the default values via engine startup parameters. For implementation details, refer to: [vLLM](https://github.com/vllm-project/vllm/pull/34330) / [SGLang](https://github.com/sgl-project/sglang/pull/18467).

1051

1052

1053

### Citation

1054

1055

If you find our work helpful, feel free to give us a cite.

1056

1057

```bibtex

1058

@misc{qwen3.6-27b,

1059

title = {{Qwen3.6-27B}: Flagship-Level Coding in a {27B} Dense Model},

1060

author = {{Qwen Team}},

1061

month = {April},

1062

year = {2026},

1063

url = {https://qwen.ai/blog?id=qwen3.6-27b}

1064

}

1065

```