README.md
60.3 KB · 970 lines · markdown Raw
1 ---
2 library_name: transformers
3 license: apache-2.0
4 license_link: https://huggingface.co/Qwen/Qwen3.5-0.8B/blob/main/LICENSE
5 pipeline_tag: image-text-to-text
6 base_model:
7 - Qwen/Qwen3.5-0.8B-Base
8 ---
9
10 # Qwen3.5-0.8B
11
12 <img width="400px" src="https://qianwen-res.oss-accelerate.aliyuncs.com/logo_qwen3.5.png">
13
14 [![Qwen Chat](https://img.shields.io/badge/%F0%9F%92%9C%EF%B8%8F%20Qwen%20Chat%20-536af5)](https://chat.qwen.ai)
15
16 > [!Note]
17 > This repository contains model weights and configuration files for the post-trained model in the Hugging Face Transformers format.
18 >
19 > These artifacts are compatible with Hugging Face Transformers, vLLM, SGLang, KTransformers, etc.
20 >
21 > In light of its parameter scale, the intended use cases are prototyping, task-specific fine-tuning, and other research or development purposes.
22
23 Over recent months, we have intensified our focus on developing foundation models that deliver exceptional utility and performance. Qwen3.5 represents a significant leap forward, integrating breakthroughs in multimodal learning, architectural efficiency, reinforcement learning scale, and global accessibility to empower developers and enterprises with unprecedented capability and efficiency.
24
25 ## Qwen3.5 Highlights
26
27 Qwen3.5 features the following enhancement:
28
29 - **Unified Vision-Language Foundation**: Early fusion training on multimodal tokens achieves cross-generational parity with Qwen3 and outperforms Qwen3-VL models across reasoning, coding, agents, and visual understanding benchmarks.
30
31 - **Efficient Hybrid Architecture**: Gated Delta Networks combined with sparse Mixture-of-Experts deliver high-throughput inference with minimal latency and cost overhead.
32
33 - **Scalable RL Generalization**: Reinforcement learning scaled across million-agent environments with progressively complex task distributions for robust real-world adaptability.
34
35 - **Global Linguistic Coverage**: Expanded support to 201 languages and dialects, enabling inclusive, worldwide deployment with nuanced cultural and regional understanding.
36
37 - **Next-Generation Training Infrastructure**: Near-100% multimodal training efficiency compared to text-only training and asynchronous RL frameworks supporting massive-scale agent scaffolds and environment orchestration.
38
39 For more details, please refer to our blog post [Qwen3.5](https://qwen.ai/blog?id=qwen3.5).
40
41
42 ## Model Overview
43
44 - Type: Causal Language Model with Vision Encoder
45 - Training Stage: Pre-training & Post-training
46 - Language Model
47 - Number of Parameters: 0.8B
48 - Hidden Dimension: 1024
49 - Token Embedding: 248320 (Padded)
50 - Number of Layers: 24
51 - Hidden Layout: 6 × (3 × (Gated DeltaNet → FFN) → 1 × (Gated Attention → FFN))
52 - Gated DeltaNet:
53 - Number of Linear Attention Heads: 16 for V and 16 for QK
54 - Head Dimension: 128
55 - Gated Attention:
56 - Number of Attention Heads: 8 for Q and 2 for KV
57 - Head Dimension: 256
58 - Rotary Position Embedding Dimension: 64
59 - Feed Forward Network:
60 - Intermediate Dimension: 3584
61 - LM Output: 248320 (Tied to token embedding)
62 - MTP: trained with multi-steps
63 - Context Length: 262,144 natively
64
65 ## Benchmark Results
66
67 ### Language
68
69 <div style="font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,sans-serif;max-width:1000px;margin:0 auto;padding:16px 0">
70 <table style="border-collapse:collapse;font-size:13px">
71 <thead><tr>
72 <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;">Qwen3-4B-2507</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3-1.7B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.5-2B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.5-0.8B</th></tr></thead>
73 <tbody>
74 <tr><td colspan="5" 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)">Non-Thinking Mode</td></tr>
75 <tr>
76 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMLU-Pro</td>
77 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">69.6</td>
78 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">40.2</td>
79 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">55.3</td>
80 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">29.7</td>
81 </tr>
82 <tr>
83 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMLU-Redux</td>
84 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.2</td>
85 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">64.4</td>
86 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">69.2</td>
87 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">48.5</td>
88 </tr>
89 <tr>
90 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">C-Eval</td>
91 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.2</td>
92 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">61.0</td>
93 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">65.2</td>
94 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">46.4</td>
95 </tr>
96 <tr>
97 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SuperGPQA</td>
98 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">42.8</td>
99 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">21.0</td>
100 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">30.4</td>
101 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">16.9</td>
102 </tr>
103 <tr>
104 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">IFEval</td>
105 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">83.4</td>
106 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">68.2</td>
107 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">61.2</td>
108 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">52.1</td>
109 </tr>
110 <tr>
111 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMMLU</td>
112 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">64.9</td>
113 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">46.7</td>
114 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">56.9</td>
115 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">34.1</td>
116 </tr>
117 <tr><td colspan="5" 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 & STEM (Thinking)</td></tr>
118 <tr>
119 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMLU-Pro</td>
120 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">74.0</td>
121 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">56.5</td>
122 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">66.5</td>
123 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">42.3</td>
124 </tr>
125 <tr>
126 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMLU-Redux</td>
127 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.1</td>
128 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">73.9</td>
129 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">79.6</td>
130 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">59.5</td>
131 </tr>
132 <tr>
133 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">C-Eval</td>
134 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">82.2</td>
135 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">68.1</td>
136 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">73.2</td>
137 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">50.5</td>
138 </tr>
139 <tr>
140 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SuperGPQA</td>
141 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">47.8</td>
142 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">31.2</td>
143 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">37.5</td>
144 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">21.3</td>
145 </tr>
146 <tr>
147 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">GPQA</td>
148 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">65.8</td>
149 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">40.1</td>
150 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">51.6</td>
151 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">11.9</td>
152 </tr>
153 <tr><td colspan="5" 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)">Instruction Following (Thinking）</td></tr>
154 <tr>
155 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">IFEval</td>
156 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">87.4</td>
157 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">72.5</td>
158 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">78.6</td>
159 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">44.0</td>
160 </tr>
161 <tr>
162 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">IFBench</td>
163 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">50.4</td>
164 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">26.7</td>
165 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">41.3</td>
166 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">21.0</td>
167 </tr>
168 <tr>
169 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MultiChallenge</td>
170 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">41.7</td>
171 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">27.2</td>
172 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">33.7</td>
173 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">18.9</td>
174 </tr>
175 <tr><td colspan="5" 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)">Long Context (Thinking）</td></tr>
176 <tr>
177 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">AA-LCR</td>
178 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">32.0</td>
179 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">6.7</td>
180 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">25.6</td>
181 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">4.7</td>
182 </tr>
183 <tr>
184 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">LongBench v2</td>
185 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">42.8</td>
186 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">26.5</td>
187 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">38.7</td>
188 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">26.1</td>
189 </tr>
190 <tr><td colspan="5" 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)">Reasoning (Thinking）</td></tr>
191 <tr>
192 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">HMMT Feb 25</td>
193 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">57.5</td>
194 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">10.2</td>
195 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">22.9</td>
196 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
197 </tr>
198 <tr>
199 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">HMMT Nov 25</td>
200 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">69.6</td>
201 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">8.9</td>
202 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">19.6</td>
203 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
204 </tr>
205 <tr><td colspan="5" 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 Agent (Thinking）</td></tr>
206 <tr>
207 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">BFCL-V4</td>
208 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">39.9</td>
209 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
210 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">43.6</td>
211 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">25.3</td>
212 </tr>
213 <tr>
214 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">TAU2-Bench</td>
215 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">43.2</td>
216 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--</td>
217 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">48.8</td>
218 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">11.6</td>
219 </tr>
220 <tr><td colspan="5" 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)">Multilingualism (Thinking）</td></tr>
221 <tr>
222 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMMLU</td>
223 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">70.8</td>
224 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">57.0</td>
225 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">63.1</td>
226 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">44.3</td>
227 </tr>
228 <tr>
229 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMLU-ProX</td>
230 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">62.4</td>
231 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">49.4</td>
232 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">52.3</td>
233 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">34.6</td>
234 </tr>
235 <tr>
236 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">NOVA-63</td>
237 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">47.1</td>
238 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">40.3</td>
239 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">46.4</td>
240 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">42.4</td>
241 </tr>
242 <tr>
243 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">INCLUDE</td>
244 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">64.4</td>
245 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">51.8</td>
246 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">55.4</td>
247 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">40.6</td>
248 </tr>
249 <tr>
250 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">Global PIQA</td>
251 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">73.5</td>
252 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">63.1</td>
253 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">69.3</td>
254 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">59.4</td>
255 </tr>
256 <tr>
257 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">PolyMATH</td>
258 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">46.2</td>
259 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">25.2</td>
260 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">26.1</td>
261 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">8.2</td>
262 </tr>
263 <tr>
264 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">WMT24++</td>
265 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">58.9</td>
266 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">39.3</td>
267 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">45.8</td>
268 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">27.2</td>
269 </tr>
270 <tr>
271 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MAXIFE</td>
272 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">72.1</td>
273 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">50.7</td>
274 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">60.6</td>
275 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">39.2</td>
276 </tr>
277 </tbody>
278 </table>
279 <p style="margin-top:12px;font-size:11px;opacity:0.7">
280 * TAU2-Bench: we follow the official setup except for the airline domain, where all models are evaluated by applying the fixes proposed in the Claude Opus 4.5 system card.
281 <br>
282 * MMLU-ProX: we report the averaged accuracy on 29 languages.<br>
283 * WMT24++: a harder subset of WMT24 after difficulty labeling and rebalancing; we report the averaged scores on 55 languages using XCOMET-XXL.<br>
284 * MAXIFE: we report the accuracy on English + multilingual original prompts (totally 23 settings).<br>
285 * Experimental settings: top_p=0.95, top_k=20, presence_penalty=1.5, and temperature=1.0 were used.<br>
286 * Empty cells (--) indicate scores not yet available or not applicable.
287 </p>
288 </div>
289
290 ### Vision Language
291
292
293 <div style="font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,sans-serif;max-width:1000px;margin:0 auto;padding:16px 0">
294 <table style="width:100%;border-collapse:collapse;font-size:13px">
295 <thead><tr>
296 <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-VL-4B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3-VL-2B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.5-2B</th><th style="padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #7c3aed;color:#7c3aed;font-size: 14px;">Qwen3.5-0.8B</th></tr></thead>
297 <tbody>
298 <tr><td colspan="5" 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 and Puzzle</td></tr>
299 <tr>
300 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMMU</td>
301 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">70.8</td>
302 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">61.4</td>
303 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">64.2/64.2</td>
304 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">49/47.4</td>
305 </tr>
306 <tr>
307 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMMU-Pro</td>
308 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">57.0</td>
309 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">42.5</td>
310 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">50.3/47.7</td>
311 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">31.2/31.4</td>
312 </tr>
313 <tr>
314 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">Mathvista(mini)</td>
315 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">79.5</td>
316 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">73.6</td>
317 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">76.7/73.9</td>
318 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">62.2/58.6</td>
319 </tr>
320 <tr>
321 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">DynaMath</td>
322 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">74.4</td>
323 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">66.7</td>
324 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">73.6/69.6</td>
325 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">49.9/46.5</td>
326 </tr>
327 <tr>
328 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">ZEROBench</td>
329 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">0.0</td>
330 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">0.0</td>
331 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">1.0/0.0</td>
332 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">0.0/0.0</td>
333 </tr>
334 <tr>
335 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">ZEROBench_sub</td>
336 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">18.9</td>
337 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">13.2</td>
338 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">17.1/18.6</td>
339 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">12.9/11.4</td>
340 </tr>
341 <tr>
342 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">VlmsAreBlind</td>
343 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">68.6</td>
344 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">50.0</td>
345 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">75.8/74.3</td>
346 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">59.4/57.3</td>
347 </tr>
348 <tr><td colspan="5" 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>
349 <tr>
350 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">RealWorldQA</td>
351 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">73.2</td>
352 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">69.5</td>
353 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">74.5/71.2</td>
354 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">63.4/61.6</td>
355 </tr>
356 <tr>
357 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMStar</td>
358 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">73.2</td>
359 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">68.1</td>
360 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">71.7/68.0</td>
361 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">58.3/55.9</td>
362 </tr>
363 <tr>
364 <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>
365 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.7</td>
366 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">81.9</td>
367 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">83.3/81.3</td>
368 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">69.9/68.0</td>
369 </tr>
370 <tr>
371 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SimpleVQA</td>
372 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">48.8</td>
373 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">43.6</td>
374 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">38.5/39.5</td>
375 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">31.3/30.4</td>
376 </tr>
377 <tr>
378 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">HallusionBench</td>
379 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">64.1</td>
380 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">54.9</td>
381 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">58.0/51.3</td>
382 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">53.1/46.7</td>
383 </tr>
384 <tr><td colspan="5" 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)">Text Recognition and Document Understanding</td></tr>
385 <tr>
386 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMLongBench-Doc</td>
387 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">44.4</td>
388 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">33.8</td>
389 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">45.4/38.8</td>
390 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">33.6/28.1</td>
391 </tr>
392 <tr>
393 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">AI2D_TEST</td>
394 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.9</td>
395 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.4</td>
396 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">83.3/81.5</td>
397 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">69.9/68.7</td>
398 </tr>
399 <tr>
400 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">CC-OCR</td>
401 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">73.8</td>
402 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">68.3</td>
403 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">72.9/75.8</td>
404 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">63.2/66.7</td>
405 </tr>
406 <tr>
407 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">OmniDocBench1.5</td>
408 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.0</td>
409 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">65.9</td>
410 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">79.8/80.9</td>
411 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">61.0/70.6</td>
412 </tr>
413 <tr>
414 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">CharXiv(RQ)</td>
415 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">50.3</td>
416 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">37.1</td>
417 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">58.8/52.6</td>
418 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">41.3/38.2</td>
419 </tr>
420 <tr>
421 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">OCRBench</td>
422 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.8</td>
423 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">79.2</td>
424 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.5/85.4</td>
425 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">74.5/79.1</td>
426 </tr>
427 <tr><td colspan="5" 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>
428 <tr>
429 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">RefCOCO(avg)</td>
430 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">88.2</td>
431 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.8</td>
432 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.8/84.3</td>
433 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">79.3/77.8</td>
434 </tr>
435 <tr>
436 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">CountBench</td>
437 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">89.4</td>
438 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.1</td>
439 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">91.4/86.8</td>
440 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">77.0/68.6</td>
441 </tr>
442 <tr>
443 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">ODInW13</td>
444 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">39.4</td>
445 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">36.0</td>
446 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">35.9/40.5</td>
447 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">31.6/33.2</td>
448 </tr>
449 <tr>
450 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">ERQA</td>
451 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">47.3</td>
452 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">41.8</td>
453 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">43.8/33.0</td>
454 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">34.5/23.8</td>
455 </tr>
456 <tr>
457 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">EmbSpatialBench</td>
458 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">80.7</td>
459 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">75.9</td>
460 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">77.9/66.4</td>
461 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">68.6/54.6</td>
462 </tr>
463 <tr>
464 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">RefSpatialBench</td>
465 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">45.3</td>
466 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">28.9</td>
467 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">32.9/30.0</td>
468 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">23.5/21.7</td>
469 </tr>
470 <tr>
471 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">Hypersim</td>
472 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">11.9</td>
473 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">11.2</td>
474 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">12.4/12.4</td>
475 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">11.9/11.0</td>
476 </tr>
477 <tr>
478 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SUNRGBD</td>
479 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">28.0</td>
480 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">28.6</td>
481 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">28.7/25.6</td>
482 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">26.1/23.3</td>
483 </tr>
484 <tr>
485 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">Nuscene</td>
486 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">4.9</td>
487 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">4.0</td>
488 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">6.9/8.5</td>
489 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">5.7/7.0</td>
490 </tr>
491 <tr><td colspan="5" 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>
492 <tr>
493 <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>
494 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">76.0</td>
495 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">67.9</td>
496 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">75.6/--</td>
497 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">63.8/--</td>
498 </tr>
499 <tr>
500 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">VideoMME<sub><small>(w/o sub.)</sub></small></td>
501 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">68.9</td>
502 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">62.1</td>
503 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">69.0/--</td>
504 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">57.7/--</td>
505 </tr>
506 <tr>
507 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">VideoMMMU</td>
508 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">69.4</td>
509 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">54.1</td>
510 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">62.1/--</td>
511 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">44.3/--</td>
512 </tr>
513 <tr>
514 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MLVU</td>
515 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">75.7</td>
516 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">69.2</td>
517 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">76.2/--</td>
518 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">65.6/--</td>
519 </tr>
520 <tr>
521 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MVBench</td>
522 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">69.3</td>
523 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">64.5</td>
524 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">64.9/--</td>
525 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">55.8/--</td>
526 </tr>
527 <tr>
528 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">LVBench</td>
529 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">53.5</td>
530 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">47.6</td>
531 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">57.1/--</td>
532 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">45.1/--</td>
533 </tr>
534 <tr>
535 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MMVU</td>
536 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">58.6</td>
537 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">48.9</td>
538 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">48.6/--</td>
539 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">34.3/--</td>
540 </tr>
541 <tr><td colspan="5" 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>
542 <tr>
543 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">ScreenSpot Pro</td>
544 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">59.5</td>
545 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">48.5</td>
546 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--/54.5</td>
547 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">--/46.5</td>
548 </tr>
549 <tr><td colspan="5" 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)">Medical VQA</td></tr>
550 <tr>
551 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SLAKE</td>
552 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">65.9</td>
553 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">61.1</td>
554 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">74.4/67.5</td>
555 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">62.6/59.5</td>
556 </tr>
557 <tr>
558 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">PMC-VQA</td>
559 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">48.4</td>
560 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">42.4</td>
561 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">48.8/54.0</td>
562 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">40.4/45.5</td>
563 </tr>
564 <tr>
565 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MedXpertQA-MM</td>
566 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">26.3</td>
567 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">13.0</td>
568 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">26.9/19.1</td>
569 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">17.1/25.3</td>
570 </tr>
571 </tbody>
572 </table>
573
574 <p style="margin-top:12px;font-size:11px;opacity:0.7">
575 * Scores of Qwen3.5 models are reported as Thinking / Non-thinking.<br>
576 * MathVision: our model’s score is evaluated using a fixed prompt, e.g., “Please reason step by step, and put your final answer within \boxed{}.” For other models, we report the higher score between runs with and without the \boxed{} formatting.<br>
577 * Experimental settings: For the Video benchmarks, we used top_p=0.95, top_k=20, presence_penalty=1.5, and temperature=1.0. All other benchmarks adopted the same sampling configuration but with temperature=0.6 under the thinking mode. Under the non-thinking mode, the sampling parameters were set to top_p=0.8, top_k=20, presence_penalty=1.5, and temperature=0.7.<br>
578 * Empty cells (--) indicate scores not yet available or not applicable.
579 </p>
580 </div>
581
582 ## Quickstart
583
584 > [!Important]
585 > Qwen3.5 models support both non-thinking and thinking mode. **Qwen3.5-0.8B operates in non-thinking mode by default**.
586 > To enable thinking, refer to the examples [here](#thinking-mode).
587
588 For streamlined integration, we recommend using Qwen3.5 via APIs. Below is a guide to use Qwen3.5 via OpenAI-compatible API.
589
590 ### Serving Qwen3.5
591
592 Qwen3.5 can be served via APIs with popular inference frameworks.
593 In the following, we show example commands to launch OpenAI-Compatible API servers for Qwen3.5 models.
594
595 > [!Important]
596 > Inference efficiency and throughput vary significantly across frameworks.
597 > We recommend using the latest framework versions to ensure optimal performance and compatibility.
598 > For production workloads or high-throughput scenarios, dedicated serving engines such as SGLang, KTransformers or vLLM are strongly recommended.
599
600 > [!Important]
601 > The model has a default context length of 262,144 tokens.
602 > If you encounter out-of-memory (OOM) errors, consider reducing the context window.
603
604 #### SGLang
605
606 [SGLang](https://github.com/sgl-project/sglang) is a fast serving framework for large language models and vision language models.
607 SGLang from the main branch of the open-source repository is required for Qwen3.5, which can be installed using the following command in a fresh environment:
608 ```shell
609 uv pip install 'git+https://github.com/sgl-project/sglang.git#subdirectory=python&egg=sglang[all]'
610 ```
611 See [its documentation](https://docs.sglang.ai/get_started/install.html) for more details.
612
613 The following will create API endpoints at `http://localhost:8000/v1`:
614
615 - **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.
616
617 ```shell
618 python -m sglang.launch_server --model-path Qwen/Qwen3.5-0.8B --port 8000 --tp-size 1 --mem-fraction-static 0.8 --context-length 262144
619 ```
620
621 - **Tool Use**: To support tool use, you can use the following command.
622
623 ```shell
624 python -m sglang.launch_server --model-path Qwen/Qwen3.5-0.8B --port 8000 --tp-size 1 --mem-fraction-static 0.8 --context-length 262144 --tool-call-parser qwen3_coder
625 ```
626
627 - **Multi-Token Prediction (MTP)**: The following command is recommended for MTP:
628
629 ```shell
630 python -m sglang.launch_server --model-path Qwen/Qwen3.5-0.8B --port 8000 --tp-size 1 --mem-fraction-static 0.8 --context-length 262144 --speculative-algo NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4
631 ```
632
633 #### vLLM
634
635 [vLLM](https://github.com/vllm-project/vllm) is a high-throughput and memory-efficient inference and serving engine for LLMs.
636 vLLM from the main branch of the open-source repository is required for Qwen3.5, which can be installed using the following command in a fresh environment:
637 ```shell
638 uv pip install vllm --torch-backend=auto --extra-index-url https://wheels.vllm.ai/nightly
639 ```
640 See [its documentation](https://docs.vllm.ai/en/stable/getting_started/installation/index.html) for more details.
641
642 For detailed Qwen3.5 usage guide, see the [vLLM Qwen3.5 recipe](https://docs.vllm.ai/projects/recipes/en/latest/Qwen/Qwen3.5.html).
643
644 The following will create API endpoints at `http://localhost:8000/v1`:
645
646 - **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.
647
648 ```shell
649 vllm serve Qwen/Qwen3.5-0.8B --port 8000 --tensor-parallel-size 1 --max-model-len 262144
650 ```
651
652 - **Tool Call**: To support tool use, you can use the following command.
653
654 ```shell
655 vllm serve Qwen/Qwen3.5-0.8B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --enable-auto-tool-choice --tool-call-parser qwen3_coder
656 ```
657
658 - **Multi-Token Prediction (MTP)**: The following command is recommended for MTP:
659
660 ```shell
661 vllm serve Qwen/Qwen3.5-0.8B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --speculative-config '{"method":"qwen3_next_mtp","num_speculative_tokens":2}'
662 ```
663
664 - **Text-Only**: The following command skips the vision encoder and multimodal profiling to free up memory for additional KV cache:
665
666 ```shell
667 vllm serve Qwen/Qwen3.5-0.8B --port 8000 --tensor-parallel-size 1 --max-model-len 262144 --language-model-only
668 ```
669
670 #### KTransformers
671
672 [KTransformers](https://github.com/kvcache-ai/ktransformers) is a flexible framework for experiencing cutting-edge LLM inference optimizations with CPU-GPU heterogeneous computing.
673 For running Qwen3.5 with KTransformers, see the [KTransformers Deployment Guide](https://github.com/kvcache-ai/ktransformers/blob/main/doc/en/Qwen3.5.md).
674
675 #### Hugging Face Transformers
676
677 Hugging Face Transformers contains a _lightweight_ server which can be used for quick testing and moderate load deployment.
678 The latest `transformers` is required for Qwen3.5:
679 ```shell
680 pip install "transformers[serving] @ git+https://github.com/huggingface/transformers.git@main"
681 ```
682 See [its documentation](https://huggingface.co/docs/transformers/main/serving) for more details. Please also make sure torchvision and pillow are installed.
683
684 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:
685 ```shell
686 transformers serve --force-model Qwen/Qwen3.5-0.8B --port 8000 --continuous-batching
687 ```
688
689 ### Using Qwen3.5 via the Chat Completions API
690
691 The chat completions API is accessible via standard HTTP requests or OpenAI SDKs.
692 Here, we show examples using the OpenAI Python SDK.
693
694 Before starting, make sure it is installed and the API key and the API base URL is configured, e.g.:
695 ```shell
696 pip install -U openai
697
698 # Set the following accordingly
699 export OPENAI_BASE_URL="http://localhost:8000/v1"
700 export OPENAI_API_KEY="EMPTY"
701 ```
702
703 > [!Tip]
704 > We recommend using the following set of sampling parameters for generation
705 > - Non-thinking mode for text tasks: `temperature=1.0, top_p=1.00, top_k=20, min_p=0.0, presence_penalty=2.0, repetition_penalty=1.0`
706 > - Non-thinking mode for VL tasks: `temperature=0.7, top_p=0.80, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0`
707 > - Thinking mode for text tasks: `temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0`
708 > - Thinking mode for VL or precise coding (e.g. WebDev) tasks : `temperature=0.6, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0`
709 >
710 > Please note that the support for sampling parameters varies according to inference frameworks.
711
712 #### Text-Only Input
713
714 ```python
715 from openai import OpenAI
716 # Configured by environment variables
717 client = OpenAI()
718
719 messages = [
720 {"role": "user", "content": "Give me a short introduction to large language models."},
721 ]
722
723 chat_response = client.chat.completions.create(
724 model="Qwen/Qwen3.5-0.8B",
725 messages=messages,
726 max_tokens=32768,
727 temperature=1.0,
728 top_p=1.0,
729 presence_penalty=2.0,
730 extra_body={
731 "top_k": 20,
732 },
733 )
734 print("Chat response:", chat_response)
735 ```
736
737 #### Image Input
738
739 ```python
740 from openai import OpenAI
741 # Configured by environment variables
742 client = OpenAI()
743
744 messages = [
745 {
746 "role": "user",
747 "content": [
748 {
749 "type": "image_url",
750 "image_url": {
751 "url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/RealWorld/RealWorld-04.png"
752 }
753 },
754 {
755 "type": "text",
756 "text": "Where is this?"
757 }
758 ]
759 }
760 ]
761
762 chat_response = client.chat.completions.create(
763 model="Qwen/Qwen3.5-0.8B",
764 messages=messages,
765 max_tokens=32768,
766 temperature=0.7,
767 top_p=0.8,
768 presence_penalty=1.5,
769 extra_body={
770 "top_k": 20,
771 },
772 )
773 print("Chat response:", chat_response)
774 ```
775
776 #### Video Input
777
778 ```python
779 from openai import OpenAI
780 # Configured by environment variables
781 client = OpenAI()
782
783 messages = [
784 {
785 "role": "user",
786 "content": [
787 {
788 "type": "video_url",
789 "video_url": {
790 "url": "https://qianwen-res.oss-accelerate.aliyuncs.com/Qwen3.5/demo/video/N1cdUjctpG8.mp4"
791 }
792 },
793 {
794 "type": "text",
795 "text": "Summarize the video content."
796 }
797 ]
798 }
799 ]
800
801 # When vLLM is launched with `--media-io-kwargs '{"video": {"num_frames": -1}}'`,
802 # video frame sampling can be configured via `extra_body` (e.g., by setting `fps`).
803 # This feature is currently supported only in vLLM.
804 #
805 # By default, `fps=2` and `do_sample_frames=True`.
806 # With `do_sample_frames=True`, you can customize the `fps` value to set your desired video sampling rate.
807 chat_response = client.chat.completions.create(
808 model="Qwen/Qwen3.5-0.8B",
809 messages=messages,
810 max_tokens=32768,
811 temperature=0.7,
812 top_p=0.8,
813 presence_penalty=1.5,
814 extra_body={
815 "top_k": 20,
816 "mm_processor_kwargs": {"fps": 2, "do_sample_frames": True},
817 },
818 )
819
820 print("Chat response:", chat_response)
821 ```
822
823 #### Thinking Mode
824
825 > [!Important]
826 > Qwen3.5 does not officially support the soft switch of Qwen3, i.e., `/think` and `/nothink`.
827
828 You can make the model think before response by configuring the API parameters.
829 For example,
830
831 ```python
832 from openai import OpenAI
833 # Configured by environment variables
834 client = OpenAI()
835
836 messages = [
837 {"role": "user", "content": "Type \"I love Qwen3.5\" backwards"},
838 ]
839
840 chat_response = client.chat.completions.create(
841 model="Qwen/Qwen3.5-0.8B",
842 messages=messages,
843 max_tokens=81920,
844 temperature=1.0,
845 top_p=0.95,
846 presence_penalty=1.5,
847 extra_body={
848 "top_k": 20,
849 "enable_thinking": True,
850 },
851 )
852 print("Chat response:", chat_response)
853 ```
854
855 > [!Important]
856 > In thinking mode, we have observed that when using the recommended sampling parameters, Qwen3.5-0.8B is more prone to entering thinking loops compared to other Qwen3.5 models, which may prevent it from terminating generation properly.
857 > We recommend further tuning the sampling parameters specific to your use case and utilizing the API's streaming generation mode (if supported) to enable timely detection and interruption of such anomalous generation behaviors.
858
859
860 ## Agentic Usage
861
862 Qwen3.5 excels in tool calling capabilities.
863
864 ### Qwen-Agent
865
866 We recommend using [Qwen-Agent](https://github.com/QwenLM/Qwen-Agent) to quickly build Agent applications with Qwen3.5.
867
868 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.
869 ```python
870 import os
871 from qwen_agent.agents import Assistant
872
873 # Define LLM
874 # Using OpenAI-compatible API endpoint. The API backend should disable response parsers.
875 llm_cfg = {
876 # Use your own model service compatible with OpenAI API by vLLM/SGLang:
877 'model': 'Qwen/Qwen3.5-0.8B',
878 'model_type': 'qwenvl_oai',
879 'model_server': 'http://localhost:8000/v1', # api_base
880 'api_key': 'EMPTY',
881
882 'generate_cfg': {
883 'use_raw_api': True,
884 # Pass the parameter of whether to enable thinking mode in this way
885 # 'extra_body': {
886 # 'chat_template_kwargs': {'enable_thinking': True}
887 # },
888 },
889 }
890
891 # Define Tools
892 tools = [
893 {'mcpServers': { # You can specify the MCP configuration file
894 "filesystem": {
895 "command": "npx",
896 "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/xxxx/Desktop"]
897 }
898 }
899 }
900 ]
901
902 # Define Agent
903 bot = Assistant(llm=llm_cfg, function_list=tools)
904
905 # Streaming generation
906 messages = [{'role': 'user', 'content': 'Help me organize my desktop.'}]
907 for responses in bot.run(messages=messages):
908 pass
909 print(responses)
910
911 # Streaming generation
912 messages = [{'role': 'user', 'content': 'Develop a dog website and save it on the desktop'}]
913 for responses in bot.run(messages=messages):
914 pass
915 print(responses)
916 ```
917
918 ### Qwen Code
919
920
921 [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.
922
923 For more information, please refer to [Qwen Code](https://qwenlm.github.io/qwen-code-docs/).
924
925 ## Best Practices
926
927 To achieve optimal performance, we recommend the following settings:
928
929 1. **Sampling Parameters**:
930 - We suggest using the following sets of sampling parameters depending on the mode and task type:
931 - **Non-thinking mode for text tasks**:
932 `temperature=1.0`, `top_p=1.00`, `top_k=20`, `min_p=0.0`, `presence_penalty=2.0`, `repetition_penalty=1.0`
933 - **Non-thinking mode for VL tasks**:
934 `temperature=0.7`, `top_p=0.80`, `top_k=20`, `min_p=0.0`, `presence_penalty=1.5`, `repetition_penalty=1.0`
935 - **Thinking mode for text tasks**:
936 `temperature=1.0`, `top_p=0.95`, `top_k=20`, `min_p=0.0`, `presence_penalty=1.5`, `repetition_penalty=1.0`
937 - **Thinking mode for VL or precise coding (e.g., WebDev) tasks**:
938 `temperature=0.6`, `top_p=0.95`, `top_k=20`, `min_p=0.0`, `presence_penalty=0.0`, `repetition_penalty=1.0`
939
940 - 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.
941
942 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.
943
944 3. **Standardize Output Format**: We recommend using prompts to standardize model outputs when benchmarking.
945 - **Math Problems**: Include "Please reason step by step, and put your final answer within \boxed{}." in the prompt.
946 - **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"`."
947
948 4. **No Thinking Content in History**: In multi-turn conversations, the historical model output should only include the final output part and does not need to include the thinking content. It is implemented in the provided chat template in Jinja2. However, for frameworks that do not directly use the Jinja2 chat template, it is up to the developers to ensure that the best practice is followed.
949
950 5. **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,
951 ```json
952 {"longest_edge": 469762048, "shortest_edge": 4096}
953 ```
954
955 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).
956
957
958 ### Citation
959
960 If you find our work helpful, feel free to give us a cite.
961
962 ```bibtex
963 @misc{qwen3.5,
964 title = {{Qwen3.5}: Towards Native Multimodal Agents},
965 author = {{Qwen Team}},
966 month = {February},
967 year = {2026},
968 url = {https://qwen.ai/blog?id=qwen3.5}
969 }
970 ```