README.md
29.2 KB · 463 lines · markdown Raw
1 ---
2 library_name: transformers
3 license: mit
4 license_link: https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B/blob/main/LICENSE
5 pipeline_tag: text-generation
6 ---
7
8
9 <img width="600px" src="assets/ornith_logo.png">
10
11 [![Ornith Blog](https://img.shields.io/badge/%F0%9F%A6%A2%EF%B8%8F%20Ornith%20Blog%20-FD8E5B)](https://deep-reinforce.com/ornith.html)
12
13
14 # Ornith-1.5-35B-A3B
15
16
17 Chirp Chirp! 🐦 We are introducing Ornith-1.5, a major step toward building foundation models through end-to-end self-improvement.
18
19 Ornith-1.5 extends Ornith-1.0 (which was developed on top of Qwen3.5 and Gemma4 with additional continued pretraining, mid-training, and post-training) by expanding the self-improvement loop from scaffold and rollout optimization to jointly optimizing task generation, scaffold construction, and solution rollouts. Rather than relying on a fixed set of human-curated tasks and manually designed harnesses, Ornith-1.5 continuously generates new training tasks, discovers effective strategies for solving them, and improves the policy through reinforcement learning. For more details on the task, harness, and rollout reward design, please refer to our [blog](https://ornith.ai/ornith_1_5.html).
20 <img style="width: 100%; max-width: 900px;" src="assets/ornith_35b_eval.png" alt="Ornith 1.5 35B Benchmark Results" title="Ornith 1.5 35B Benchmark Results">
21
22 ## Ornith 1.5 35B-A3B
23
24 This model card documents **Ornith-1.5-35B-A3B**, the mid-size mixture-of-experts member of the Ornith-1.5 family. It activates only ~3B parameters per token, yet significantly outperforms its similar-sized peer Qwen 3.6-35B across all coding and agentic benchmarks, and outperforms dense models such as Gemma 4-31B and Muse Glimmer-30B by wide margins on agentic coding.
25
26 ### Benchmarks
27
28
29 <div style="font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,sans-serif;width:100%;margin:0 auto;padding:16px 0">
30 <table style="width:100%;table-layout:fixed;border-collapse:collapse;font-size:13px">
31 <thead><tr>
32 <th style="width:22%;padding:10px 7px;text-align:left;font-weight:600;border-bottom:2px solid #FD8E5B;color:#FD8E5B"></th>
33 <th style="width:13%;padding:10px 7px;text-align:center;font-weight:700;border-bottom:2px solid #FD8E5B;color:#FD8E5B;font-size:14px;background:rgba(253, 142, 91, 0.12)">Ornith-1.5-35B-A3B</th>
34 <th style="width:13%;padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #FD8E5B;color:#FD8E5B;font-size:14px">Ornith-1.0-35B-A3B</th>
35 <th style="width:13%;padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #FD8E5B;color:#FD8E5B;font-size:14px">Qwen3.6-35B-A3B</th>
36 <th style="width:13%;padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #FD8E5B;color:#FD8E5B;font-size:14px">Gemma-4-31B <sub></sub></th>
37 <th style="width:13%;padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #FD8E5B;color:#FD8E5B;font-size:14px">Muse-Glimmer-30B <sub></sub></th>
38 <th style="width:13%;padding:10px 7px;text-align:center;font-weight:500;border-bottom:2px solid #FD8E5B;color:#FD8E5B;font-size:14px">Qwen3.5-397B</th>
39 </tr></thead>
40 <tbody>
41 <tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#FD8E5B;border-bottom:1px solid rgba(253, 142, 91, 0.2);background:rgba(253, 142, 91, 0.1)">Coding</td></tr>
42 <tr>
43 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">Terminal-Bench 2.1 <sub><small>(Terminus-2)</small></sub></td>
44 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15);font-weight:600;color:#FD8E5B;background:rgba(253, 142, 91, 0.06)">67.8</td>
45 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">64.2</td>
46 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">52.5</td>
47 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">42.1</td>
48 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">51.7</td>
49 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">53.5</td>
50 </tr>
51 <tr>
52 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">Terminal-Bench 2.1 <sub><small>(Claude Code)</small></sub></td>
53 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15);font-weight:600;color:#FD8E5B;background:rgba(253, 142, 91, 0.06)">68.5</td>
54 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">62.8</td>
55 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">49.2</td>
56 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">-</td>
57 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">-</td>
58 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">48.6</td>
59 </tr>
60 <tr>
61 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SWE-bench Verified</td>
62 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15);font-weight:600;color:#FD8E5B;background:rgba(253, 142, 91, 0.06)">79</td>
63 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">75.6</td>
64 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">73.4</td>
65 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">52</td>
66 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">76</td>
67 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">76.4</td>
68 </tr>
69 <tr>
70 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SWE-bench Pro</td>
71 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15);font-weight:600;color:#FD8E5B;background:rgba(253, 142, 91, 0.06)">59.6</td>
72 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">50.4</td>
73 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">49.5</td>
74 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">35.7</td>
75 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">51.2</td>
76 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">51.6</td>
77 </tr>
78 <tr>
79 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SWE-bench Multilingual</td>
80 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15);font-weight:600;color:#FD8E5B;background:rgba(253, 142, 91, 0.06)">71.4</td>
81 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">69.3</td>
82 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">67.2</td>
83 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">51.7</td>
84 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">-</td>
85 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">69.3</td>
86 </tr>
87 <tr>
88 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">DeepSWE</td>
89 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15);font-weight:600;color:#FD8E5B;background:rgba(253, 142, 91, 0.06)">22</td>
90 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">0</td>
91 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">0</td>
92 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">-</td>
93 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">-</td>
94 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">1</td>
95 </tr>
96 <tr>
97 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">Frontier-Bench v0.1</td>
98 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15);font-weight:600;color:#FD8E5B;background:rgba(253, 142, 91, 0.06)">5.1</td>
99 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">1.4</td>
100 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">1.4</td>
101 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">-</td>
102 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">-</td>
103 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">1.4</td>
104 </tr>
105 <tr>
106 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">NL2Repo</td>
107 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15);font-weight:600;color:#FD8E5B;background:rgba(253, 142, 91, 0.06)">46.2</td>
108 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">34.6</td>
109 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">29.4</td>
110 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">15.5</td>
111 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">-</td>
112 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">36.8</td>
113 </tr>
114 <tr>
115 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">SWE Atlas - QnA</td>
116 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15);font-weight:600;color:#FD8E5B;background:rgba(253, 142, 91, 0.06)">39.8</td>
117 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">37.1</td>
118 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">15.5</td>
119 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">-</td>
120 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">-</td>
121 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">20.4</td>
122 </tr>
123 <tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#FD8E5B;border-bottom:1px solid rgba(253, 142, 91, 0.2);background:rgba(253, 142, 91, 0.1)">Reasoning</td></tr>
124 <tr>
125 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">HLE <sub><small>(no tools)</small></sub></td>
126 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15);font-weight:600;color:#FD8E5B;background:rgba(253, 142, 91, 0.06)">25.6</td>
127 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">20.8</td>
128 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">21.4</td>
129 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">19.5</td>
130 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">22</td>
131 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">28.7</td>
132 </tr>
133 <tr>
134 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">HLE <sub><small>(with tools)</small></sub></td>
135 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15);font-weight:600;color:#FD8E5B;background:rgba(253, 142, 91, 0.06)">33.4</td>
136 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">30.1</td>
137 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">28.9</td>
138 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">26.5</td>
139 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">-</td>
140 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">48.3</td>
141 </tr>
142 <tr>
143 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">GPQA Diamond</td>
144 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15);font-weight:600;color:#FD8E5B;background:rgba(253, 142, 91, 0.06)">89.2</td>
145 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86.2</td>
146 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">86</td>
147 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">84.3</td>
148 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">83.5</td>
149 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">88.4</td>
150 </tr>
151 <tr><td colspan="7" style="padding:8px 12px;font-weight:600;color:#FD8E5B;border-bottom:1px solid rgba(253, 142, 91, 0.2);background:rgba(253, 142, 91, 0.1)">Agentic</td></tr>
152 <tr>
153 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">MCP-Atlas</td>
154 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15);font-weight:600;color:#FD8E5B;background:rgba(253, 142, 91, 0.06)">70.2</td>
155 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">64.4</td>
156 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">62.8</td>
157 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">55</td>
158 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">75.5</td>
159 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">72.3</td>
160 </tr>
161 <tr>
162 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">Toolathlon-Verified</td>
163 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15);font-weight:600;color:#FD8E5B;background:rgba(253, 142, 91, 0.06)">48.7</td>
164 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">42.4</td>
165 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">41.7</td>
166 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">40.8</td>
167 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">-</td>
168 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">38.3</td>
169 </tr>
170 <tr>
171 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">WideSearch</td>
172 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15);font-weight:600;color:#FD8E5B;background:rgba(253, 142, 91, 0.06)">67.8</td>
173 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">63.4</td>
174 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">60.1</td>
175 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">54.2</td>
176 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">-</td>
177 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">74</td>
178 </tr>
179 <tr>
180 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">BrowseComp</td>
181 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15);font-weight:600;color:#FD8E5B;background:rgba(253, 142, 91, 0.06)">67.6</td>
182 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">63.5</td>
183 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">62</td>
184 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">-</td>
185 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">-</td>
186 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">78.6</td>
187 </tr>
188 <tr>
189 <td style="padding:7px 7px;padding-left:20px;border-bottom:1px solid rgba(128, 128, 128, 0.15);">ClawEval</td>
190 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15);font-weight:600;color:#FD8E5B;background:rgba(253, 142, 91, 0.06)">72.5</td>
191 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">69.8</td>
192 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">68.7</td>
193 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">48.5</td>
194 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">-</td>
195 <td style="padding:7px 7px;text-align:center;border-bottom:1px solid rgba(128, 128, 128, 0.15)">70.7</td>
196 </tr>
197 </tbody>
198 </table>
199
200 <p style="margin-top:12px;font-size:10px;opacity:0.7">
201 * All results reported for Ornith-1.5 are averaged over five independent runs.<br/>
202 * Terminal-Bench 2.1 (Terminus-2): We evaluate Terminal-Bench 2.1 using the Harbor/Terminus-2 framework with parser=json, temperature=1.0, top_p=1.0, and a 128K context window. Each run uses a 4-hour timeout with 32 CPU cores and 48GB RAM, and results are averaged over 5 runs. We adjust the Qwen chat template to ensure consistency between training and inference (https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B/blob/main/chat_template.jinja), and modify Harbor to align with vLLM's reasoning_content key.<br/>
203 * Terminal-Bench 2.1 (Claude Code): We evaluate Terminal-Bench 2.1 using Claude Code 2.1.126 with parser=json, temperature=1.0, top_p=1.0, max_new_tokens=131072. Results are averaged over 5 runs. Again, Qwen chat template needs to be modified.<br/>
204 * SWE-Bench Verified, Pro and Multilingual: using OpenHands harness with temp=1.0, top_p=0.95, 256k context window. Anti-hacking safeguards are applied throughout evaluation: Git history is removed from the local repository image to prevent access to prior solutions or commits; network access is disabled, preventing the model from retrieving external information or resources.<br/>
205 * DeepSWE: Evaluated using the Claude Code harness with temperature=1.0, top_p=0.95, and a 256K context window.<br/>
206 * SWE Atlas QnA: using mini SWE agent harness with temp=1.0, top_p=0.95, 128K context window. Results are averaged over 5 runs.<br/>
207 * NL2Repo: with temperature=1.0, top_p=1.0, 400K context, 48K output. Access to specified GitHub repositories and pip packages is blocked to prevent reward hacking.<br/>
208 * HLE: Evaluated using Claude 4.6 Opus as the judge model.<br/>
209 * MCP-Atlas: All models were evaluated in thinking mode on the 500-task public subset, with a 10-minute timeout per task. We use Claude 4.8 Opus as the judge model.<br/>
210 * Toolathlon-Verified: We use the official evaluation service with the maximum token limit set to 128K.<br/>
211 * ClawEval: An agentic code benchmark over real-user task distributions; temp=0.6 and 256K context.<br/>
212 </p>
213
214 </div>
215
216
217 ## Quickstart
218
219 <div style="border-left:4px solid #FD8E5B;background:rgba(253,142,91,0.1);border-radius:6px;padding:12px 16px;font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,sans-serif;font-size:14px;line-height:1.6">
220 <div style="font-weight:700;color:#FD8E5B;margin-bottom:6px">📝 NOTE</div>
221 <p style="margin:0 0 10px"><b>Ornith-1.5-35B-A3B</b> is a <b>reasoning model</b>: by default the assistant turn opens with a <code style="background:rgba(253,142,91,0.15);padding:1px 5px;border-radius:4px">&lt;think&gt; … &lt;/think&gt;</code> block before the final answer. The serving recipes below enable a reasoning parser so the chain-of-thought is returned in a separate <code style="background:rgba(253,142,91,0.15);padding:1px 5px;border-radius:4px">reasoning_content</code> field, and a tool-call parser so the model's <code style="background:rgba(253,142,91,0.15);padding:1px 5px;border-radius:4px">&lt;tool_call&gt;</code> blocks are surfaced as OpenAI-style <code style="background:rgba(253,142,91,0.15);padding:1px 5px;border-radius:4px">tool_calls</code>.</p>
222 <p style="margin:0 0 6px">Serving Ornith-1.5-35B-A3B requires recent runtimes:</p>
223 <ul style="margin:0 0 10px;padding-left:20px">
224 <li><b>Transformers</b> ≥ 5.8.1</li>
225 <li><b>vLLM</b> ≥ 0.19.1</li>
226 <li><b>SGLang</b> ≥ 0.5.9</li>
227 </ul>
228 <p style="margin:0 0 6px">Recommended sampling parameters:</p>
229 <ul style="margin:0;padding-left:20px">
230 <li><b>For general tasks:</b> <code style="background:rgba(253,142,91,0.15);padding:1px 5px;border-radius:4px">temperature=0.6</code>, <code style="background:rgba(253,142,91,0.15);padding:1px 5px;border-radius:4px">top_p=0.95</code>, <code style="background:rgba(253,142,91,0.15);padding:1px 5px;border-radius:4px">top_k=20</code></li>
231 <li><b>To reproduce the reported benchmarks:</b> <code style="background:rgba(253,142,91,0.15);padding:1px 5px;border-radius:4px">temperature=1.0</code></li>
232 </ul>
233 </div>
234
235
236 ### Serving Ornith-1.5-35B-A3B
237
238 Ornith-1.5-35B-A3B is a ~35B mixture-of-experts model with ~3B activated parameters per token (≈70 GB in bf16). The recipes below stand up an OpenAI-compatible server on **2× 80GB GPUs** to leave headroom for the 256K context; adjust `--tensor-parallel-size` / `--tp` to match your hardware.
239
240 #### vLLM
241
242 ```bash
243 vllm serve ornith-ai/Ornith-1.5-35B-A3B \
244 --served-model-name Ornith-1.5-35B-A3B \
245 --host 0.0.0.0 --port 8000 \
246 --tensor-parallel-size 2 \
247 --max-model-len 262144 \
248 --gpu-memory-utilization 0.90 \
249 --enable-prefix-caching \
250 --enable-auto-tool-choice --tool-call-parser qwen3_xml \
251 --reasoning-parser qwen3 \
252 --trust-remote-code
253 ```
254
255 #### SGLang
256
257 ```bash
258 python -m sglang.launch_server \
259 --model-path ornith-ai/Ornith-1.5-35B-A3B \
260 --served-model-name Ornith-1.5-35B-A3B \
261 --host 0.0.0.0 --port 8000 \
262 --tp 2 \
263 --context-length 262144 \
264 --mem-fraction-static 0.85 \
265 --tool-call-parser qwen3_coder \
266 --reasoning-parser qwen3
267 ```
268
269 #### For Long-Context
270
271 Ornith-1.5-35B-A3B handles context windows of up to 262,144 tokens. When a task's combined input and output must go beyond this limit, we suggest extending the effective window with RoPE scaling — YaRN is the technique we validate against, and it is already built into both vLLM and SGLang. With a scaling factor of 4.0, the usable window grows to roughly 1M tokens.
272
273 You can turn YaRN on in either of two ways:
274
275 - **Edit the checkpoint's `config.json`.** Add a `rope_scaling` block to the model configuration:
276
277 ```json
278 {
279 "rope_scaling": {
280 "rope_type": "yarn",
281 "factor": 4.0,
282 "original_max_position_embeddings": 262144
283 }
284 }
285 ```
286
287 - **Override at launch time.** Leave the checkpoint untouched and extend the serve commands above with the equivalent flags.
288
289 vLLM:
290
291 ```bash
292 VLLM_ALLOW_LONG_MAX_MODEL_LEN=1 vllm serve ornith-ai/Ornith-1.5-35B-A3B ... --hf-overrides '{"rope_scaling": {"rope_type": "yarn", "factor": 4.0, "original_max_position_embeddings": 262144}}' --max-model-len 1000000
293 ```
294
295 SGLang:
296
297 ```bash
298 SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 python -m sglang.launch_server ... --json-model-override-args '{"rope_scaling": {"rope_type": "yarn", "factor": 4.0, "original_max_position_embeddings": 262144}}' --context-length 1000000
299 ```
300
301 <div style="border-left:4px solid #FD8E5B;background:rgba(253,142,91,0.1);border-radius:6px;padding:12px 16px;font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,sans-serif;font-size:14px;line-height:1.6">
302 <div style="font-weight:700;color:#FD8E5B;margin-bottom:6px">📝 NOTE</div>
303 <p style="margin:0">Open-source runtimes implement YaRN <i>statically</i>: the same scaling factor is applied to every request regardless of its length, which can slightly hurt quality on ordinary-length inputs. Only enable <code style="background:rgba(253,142,91,0.15);padding:1px 5px;border-radius:4px">rope_scaling</code> when your workload genuinely needs the longer window, and size <code style="background:rgba(253,142,91,0.15);padding:1px 5px;border-radius:4px">factor</code> to match it — the target window is roughly <code style="background:rgba(253,142,91,0.15);padding:1px 5px;border-radius:4px">factor</code> × 262,144, so if your requests top out around 524,288 tokens, <code style="background:rgba(253,142,91,0.15);padding:1px 5px;border-radius:4px">factor: 2.0</code> is the better setting.</p>
304 </div>
305
306 ### Using Ornith-1.5-35B-A3B via the Chat Completions API
307
308 Once a vLLM or SGLang server is running, talk to it with any OpenAI-compatible client.
309
310 #### Basic Usage
311
312 ```python
313 from openai import OpenAI
314
315 client = OpenAI(
316 base_url="http://localhost:8000/v1",
317 api_key="EMPTY", # any non-empty string works for a local server
318 )
319
320 response = client.chat.completions.create(
321 model="Ornith-1.5-35B-A3B",
322 messages=[
323 {"role": "user", "content": "Write a one-line Python lambda that squares a number."}
324 ],
325 temperature=0.6,
326 top_p=0.95,
327 max_tokens=1024,
328 )
329
330 message = response.choices[0].message
331 # reasoning_content holds the <think> trace; content holds the final answer.
332 print("reasoning:", getattr(message, "reasoning_content", None))
333 print("answer:", message.content)
334 ```
335
336 You can also stream tokens, or hand the model tools — Ornith-1.5-35B-A3B emits well-formed function calls that the server parses into the standard `tool_calls` field:
337
338 ```python
339 tools = [
340 {
341 "type": "function",
342 "function": {
343 "name": "get_weather",
344 "description": "Get the current weather for a city",
345 "parameters": {
346 "type": "object",
347 "properties": {"city": {"type": "string"}},
348 "required": ["city"],
349 },
350 },
351 }
352 ]
353
354 response = client.chat.completions.create(
355 model="Ornith-1.5-35B-A3B",
356 messages=[{"role": "user", "content": "What is the weather in Paris right now?"}],
357 tools=tools,
358 tool_choice="auto",
359 temperature=0.6,
360 max_tokens=2048,
361 )
362
363 tool_call = response.choices[0].message.tool_calls[0]
364 print(tool_call.function.name, tool_call.function.arguments)
365 # -> get_weather {"city": "Paris"}
366 ```
367
368 You can point any OpenAI-compatible SDK (Python, Node.js, etc.) or `curl` at the same `/v1/chat/completions` endpoint.
369
370 ## Agentic Usage
371
372 Ornith-1.5-35B-A3B excels in tool-calling and agentic coding. It exposes an OpenAI-compatible endpoint with tool calling and works out of the box with standard agent frameworks.
373
374 **Examples of using Ornith with agents:**
375
376
377 #### Ollama
378 ```bash
379 ollama run hf.co/ornith-ai/Ornith-1.5-35B-A3B-GGUF
380 ```
381
382
383 #### Atomic.chat
384 ```bash
385 # Atomic.chat loads a GGUF build of Ornith (ornith-ai/Ornith-1.5-35B-A3B-GGUF)
386 # through llama.cpp's OpenAI-compatible API on port 8000.
387 llama-server -hf ornith-ai/Ornith-1.5-35B-A3B-GGUF --port 8000 -c 262144
388 ```
389
390 #### llama.cpp
391 ```bash
392 # llama.cpp — serve an OpenAI-compatible API on port 8000.
393 llama-server -hf ornith-ai/Ornith-1.5-35B-A3B-GGUF --port 8000 -c 262144
394 ```
395
396 #### Hermes Agent
397 ```bash
398 # Hermes talks to any OpenAI-compatible endpoint — point it at your Ornith server.
399 export OPENAI_BASE_URL="http://localhost:8000/v1"
400 export OPENAI_API_KEY="EMPTY"
401 export MODEL="ornith-ai/Ornith-1.5-35B-A3B"
402 ```
403
404 #### OpenClaw
405
406 ```bash
407 # OpenClaw talks to any OpenAI-compatible endpoint — point it at your Ornith server.
408 export OPENAI_BASE_URL="http://localhost:8000/v1"
409 export OPENAI_API_KEY="EMPTY"
410 export OPENAI_MODEL="ornith-ai/Ornith-1.5-35B-A3B"
411 ```
412
413 #### Unsloth Studio
414
415 ```bash
416 pip install unsloth
417
418 # Load Ornith for fast local inference or fine-tuning (Python):
419 # from unsloth import FastLanguageModel
420 # model, tokenizer = FastLanguageModel.from_pretrained(
421 # "ornith-ai/Ornith-1.5-35B-A3B",
422 # max_seq_length=262144,
423 # load_in_4bit=True,
424 # )
425 ```
426
427
428 ### Coding CLIs
429
430 Ornith-1.5-35B-A3B is optimized for terminal-based coding agents. Point any OpenAI-compatible coding CLI at your Ornith-1.5-35B-A3B endpoint (set `OPENAI_BASE_URL` and `OPENAI_API_KEY`) to understand large codebases, automate tedious work, and ship faster.
431
432 #### OpenCode
433 ```bash
434 # Register your local Ornith endpoint as a provider in ~/.config/opencode/opencode.json:
435 #
436 # {
437 # "$schema": "https://opencode.ai/config.json",
438 # "provider": {
439 # "ornith": {
440 # "npm": "@ai-sdk/openai-compatible",
441 # "name": "Ornith (local)",
442 # "options": { "baseURL": "http://localhost:8000/v1", "apiKey": "EMPTY" },
443 # "models": { "ornith-ai/Ornith-1.5-35B-A3B": { "name": "Ornith-1.5-35B-A3B" } }
444 # }
445 # }
446 # }
447
448 opencode
449 ```
450
451 ### Citation
452
453 If you find our work helpful, feel free to give us a cite.
454
455 ```bibtex
456 @misc{ornith_1_5,
457 title = {{Ornith-1.5}: From Self-Scaffolding to Self-Improvement},
458 url = {https://ornith.ai/ornith_1_5.html},
459 author = {{Ornith Team}},
460 year = {2026}
461 }
462 ```
463