chen11003 commited on
Commit
c6dd802
·
verified ·
1 Parent(s): de35fdd

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +14 -12
README.md CHANGED
@@ -36,7 +36,8 @@ K2-Horizon-0.9B is the compact dense member of the K2-Horizon family: a 0.9B-cla
36
  - **Compact reasoning model.** A 0.9B-class dense model evaluated across mathematics, coding, science, and tool-use benchmarks.
37
  - **128K context.** Supports up to 131,072 tokens with YaRN RoPE scaling.
38
  - **Multi-teacher distillation.** Trained with domain teachers for math and code, STEM, and instruction following.
39
- - **Open research release.** Weights, configuration, custom architecture code, tokenizer files, and evaluation details are included.
 
40
 
41
  ## Benchmark Results
42
 
@@ -45,10 +46,10 @@ The chart at the top of this card shows K2-Horizon-0.9B against selected referen
45
  ### Full Results
46
 
47
  <!-- TABLE:START -->
48
- <div style="font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,sans-serif;margin:0 auto;padding:8px 0 16px;overflow-x:auto"><table style="display:table;width:100%;table-layout:fixed;border-collapse:collapse;font-size:12px;margin:0"><thead><tr><th style="width:30%;border-bottom:none"></th><th colspan="4" style="padding:6px 4px 2px;text-align:center;font-size:11px;font-weight:600;letter-spacing:0.04em;text-transform:uppercase;opacity:0.65;border-bottom:1px solid rgba(128,128,128,0.25)">Reference models</th></tr><tr><th style="padding:10px 6px;text-align:left;border-bottom:2px solid #2450D6"></th><th style="padding:10px 3px;text-align:center;font-weight:600;border-bottom:2px solid #2450D6;color:#2450D6;font-size:12.5px;line-height:1.2;width:17.50%;overflow-wrap:anywhere;background:rgba(36,80,214,0.08);">K2-Horizon-0.9B</th><th style="padding:10px 3px;text-align:center;font-weight:600;border-bottom:2px solid #2450D6;color:#2450D6;font-size:12.5px;line-height:1.2;width:17.50%;overflow-wrap:anywhere;">Qwen3.5-0.8B</th><th style="padding:10px 3px;text-align:center;font-weight:600;border-bottom:2px solid #2450D6;color:#2450D6;font-size:12.5px;line-height:1.2;width:17.50%;overflow-wrap:anywhere;">OpenBMB-1B</th><th style="padding:10px 3px;text-align:center;font-weight:600;border-bottom:2px solid #2450D6;color:#2450D6;font-size:12.5px;line-height:1.2;width:17.50%;overflow-wrap:anywhere;">Qwen3.5-2B</th></tr></thead><tbody><tr><td style="padding:7px 4px 7px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;font-weight:600"># Params</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);">0.9B</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">0.8B</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">1B</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">2B</td></tr><tr><td style="padding:7px 4px 7px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;font-weight:600"># Activated params</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);">0.9B</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">0.8B</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">1B</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">2B</td></tr><tr><td style="padding:7px 4px 7px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;font-weight:600">Architecture</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);">Dense</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">Dense</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">Dense</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">Dense</td></tr><tr><td colspan="5" style="padding:6px 10px;font-weight:600;font-size:12.5px;color:#2450D6;border-bottom:1px solid rgba(36,80,214,0.25);background:rgba(36,80,214,0.12)">Math</td></tr><tr><td style="padding:6px 4px 6px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle"><div style="font-size:12.5px;font-weight:600;line-height:1.2;color:inherit">AIME 2025</div><div style="margin-top:2px;font-size:10px;opacity:0.65">Competition mathematics</div></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);"><strong>41.7</strong></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">1.04</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">40.42</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">34.17</td></tr><tr><td style="padding:6px 4px 6px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle"><div style="font-size:12.5px;font-weight:600;line-height:1.2;color:inherit">AIME 2026</div><div style="margin-top:2px;font-size:10px;opacity:0.65">Competition mathematics</div></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);"><strong>48.5</strong></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">0.21</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">40.42</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">38.75</td></tr><tr><td style="padding:6px 4px 6px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle"><div style="font-size:12.5px;font-weight:600;line-height:1.2;color:inherit">HMMT Feb 2026</div><div style="margin-top:2px;font-size:10px;opacity:0.65">Competition mathematics</div></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);"><strong>25.8</strong></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">0.57</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">23.3</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">22.73</td></tr><tr><td colspan="5" style="padding:6px 10px;font-weight:600;font-size:12.5px;color:#2450D6;border-bottom:1px solid rgba(36,80,214,0.25);background:rgba(36,80,214,0.12)">Scientific Reasoning</td></tr><tr><td style="padding:6px 4px 6px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle"><div style="font-size:12.5px;font-weight:600;line-height:1.2;color:inherit">GPQA Diamond</div><div style="margin-top:2px;font-size:10px;opacity:0.65">Graduate-level science QA</div></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);">27.3</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">11.9</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">26.26</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;"><strong>54.89</strong></td></tr><tr><td colspan="5" style="padding:6px 10px;font-weight:600;font-size:12.5px;color:#2450D6;border-bottom:1px solid rgba(36,80,214,0.25);background:rgba(36,80,214,0.12)">Coding</td></tr><tr><td style="padding:6px 4px 6px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle"><div style="font-size:12.5px;font-weight:600;line-height:1.2;color:inherit">HumanEval+</div><div style="margin-top:2px;font-size:10px;opacity:0.65">Code generation</div></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);"><strong>79.9</strong></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">16.5</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">65.2</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">75.6</td></tr><tr><td style="padding:6px 4px 6px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle"><div style="font-size:12.5px;font-weight:600;line-height:1.2;color:inherit">MBPP+</div><div style="margin-top:2px;font-size:10px;opacity:0.65">Code generation</div></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);"><strong>68</strong></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">35.4</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">60.6</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">67.7</td></tr><tr><td style="padding:6px 4px 6px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle"><div style="font-size:12.5px;font-weight:600;line-height:1.2;color:inherit">LiveCodeBench v6</div><div style="margin-top:2px;font-size:10px;opacity:0.65">Competitive coding</div></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);"><strong>37.41</strong></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">6.64</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">33.52</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">29.79</td></tr><tr><td colspan="5" style="padding:6px 10px;font-weight:600;font-size:12.5px;color:#2450D6;border-bottom:1px solid rgba(36,80,214,0.25);background:rgba(36,80,214,0.12)">Agents</td></tr><tr><td style="padding:6px 4px 6px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle"><div style="font-size:12.5px;font-weight:600;line-height:1.2;color:inherit">BFCL v4</div><div style="margin-top:2px;font-size:10px;opacity:0.65">Function calling</div></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);">28</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">25.3</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">25.15</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;"><strong>43.6</strong></td></tr></tbody></table></div>
49
  <!-- TABLE:END -->
50
 
51
- Scores in %. Bold marks the best score in each row. Protocol and provenance details are in the [Technical Appendix](APPENDIX.md#evaluation).
52
 
53
  ## Quickstart
54
 
@@ -61,12 +62,12 @@ vLLM (source at [PR #53806](https://github.com/vllm-project/vllm/pull/53806), co
61
 
62
  ```shell
63
  vllm serve IFM/K2-Horizon-0.9B \
64
- --revision main \
65
- --model-impl vllm \
66
- --tensor-parallel-size 1 \
67
  --trust-remote-code \
68
  --dtype bfloat16 \
69
  --max-model-len 131072 \
 
 
 
70
  --reasoning-parser k2_horizon \
71
  --enable-auto-tool-choice \
72
  --tool-call-parser k2_horizon
@@ -75,15 +76,15 @@ vllm serve IFM/K2-Horizon-0.9B \
75
  SGLang, from a source checkout that includes [sgl-project/sglang#37654](https://github.com/sgl-project/sglang/pull/37654). This is the recipe validated in the [SGLang K2 Horizon cookbook](https://docs.sglang.io/cookbook/autoregressive/IFM/K2-Horizon):
76
 
77
  ```shell
78
- python3 -m sglang.launch_server \
79
  --model-path IFM/K2-Horizon-0.9B \
80
- --revision main \
81
  --tp 1 \
82
  --dtype bfloat16 \
83
  --attention-backend fa3 \
84
  --reasoning-parser k2_horizon \
85
- --tool-call-parser k2_horizon \
86
- --host 0.0.0.0 --port 30000
87
  ```
88
 
89
  ### API Usage
@@ -131,11 +132,11 @@ print(tokenizer.decode(outputs[0], skip_special_tokens=True))
131
  ## Best Practices
132
 
133
  1. **Reasoning effort: always `high`.** All reported results use high reasoning effort. Pass `{"chat_template_kwargs": {"reasoning_effort": "high"}}` on every request; `medium` and `low` trade accuracy for speed and are not recommended for evaluation.
134
- 2. **Sampling parameters.** `temperature=1.0`, `top_p=0.95`.
135
  3. **Output length.** Allow at least 32,768 output tokens so reasoning is never cut off. Truncated reasoning is a failed response, not a shorter one.
136
  4. **Serving.** Use the validated SGLang recipe above: BF16, TP=1, FlashAttention-3. Full recipes for every K2-Horizon size, with measured H200 latency and throughput, are in the [SGLang cookbook](https://docs.sglang.io/cookbook/autoregressive/IFM/K2-Horizon).
137
  5. **Parsers.** Enable the `k2_horizon` reasoning parser for chat, and add the `k2_horizon` tool-call parser for agent use. Leave both off for plain completion-style generation.
138
- 6. **Revisions.** `main` is the distilled release checkpoint; `mid1_75k` and `mid2_47k` preserve the context-extension stages.
139
 
140
  ## Citation
141
 
@@ -148,3 +149,4 @@ print(tokenizer.decode(outputs[0], skip_special_tokens=True))
148
  }
149
  ```
150
 
 
 
36
  - **Compact reasoning model.** A 0.9B-class dense model evaluated across mathematics, coding, science, and tool-use benchmarks.
37
  - **128K context.** Supports up to 131,072 tokens with YaRN RoPE scaling.
38
  - **Multi-teacher distillation.** Trained with domain teachers for math and code, STEM, and instruction following.
39
+ <!-- - **Intermediate checkpoints.** Intermediate checkpoints will be released so capability changes can be studied across training rather than at a single checkpoint. -->
40
+ - **Fully open.** Training data/recipe and the training code will be made public.
41
 
42
  ## Benchmark Results
43
 
 
46
  ### Full Results
47
 
48
  <!-- TABLE:START -->
49
+ <div style="font-family:-apple-system,BlinkMacSystemFont,'Segoe UI',Roboto,sans-serif;margin:0 auto;padding:8px 0 16px;overflow-x:auto"><table style="display:table;width:100%;table-layout:fixed;border-collapse:collapse;font-size:12px;margin:0"><thead><tr><th style="width:30%;border-bottom:none"></th><th colspan="4" style="padding:6px 4px 2px;text-align:center;font-size:11px;font-weight:600;letter-spacing:0.04em;text-transform:uppercase;opacity:0.65;border-bottom:1px solid rgba(128,128,128,0.25)">Reference models</th></tr><tr><th style="padding:10px 6px;text-align:left;border-bottom:2px solid #2450D6"></th><th style="padding:10px 3px;text-align:center;font-weight:600;border-bottom:2px solid #2450D6;color:#2450D6;font-size:12.5px;line-height:1.2;width:17.50%;overflow-wrap:anywhere;background:rgba(36,80,214,0.08);">K2-Horizon-0.9B</th><th style="padding:10px 3px;text-align:center;font-weight:600;border-bottom:2px solid #2450D6;color:#2450D6;font-size:12.5px;line-height:1.2;width:17.50%;overflow-wrap:anywhere;">Qwen3.5-0.8B</th><th style="padding:10px 3px;text-align:center;font-weight:600;border-bottom:2px solid #2450D6;color:#2450D6;font-size:12.5px;line-height:1.2;width:17.50%;overflow-wrap:anywhere;">OpenBMB-1B</th><th style="padding:10px 3px;text-align:center;font-weight:600;border-bottom:2px solid #2450D6;color:#2450D6;font-size:12.5px;line-height:1.2;width:17.50%;overflow-wrap:anywhere;">Qwen3.5-2B</th></tr></thead><tbody><tr><td style="padding:7px 4px 7px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;font-weight:600"># Params</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);">0.9B</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">0.8B</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">1B</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">2B</td></tr><tr><td style="padding:7px 4px 7px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;font-weight:600"># Activated params</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);">0.9B</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">0.8B</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">1B</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">2B</td></tr><tr><td style="padding:7px 4px 7px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;font-weight:600">Architecture</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);">Dense</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">Dense</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">Dense</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">Dense</td></tr><tr><td colspan="5" style="padding:6px 10px;font-weight:600;font-size:12.5px;color:#2450D6;border-bottom:1px solid rgba(36,80,214,0.25);background:rgba(36,80,214,0.12)">Math</td></tr><tr><td style="padding:6px 4px 6px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle"><div style="font-size:12.5px;font-weight:600;line-height:1.2;color:inherit">AIME 2025</div><div style="margin-top:2px;font-size:10px;opacity:0.65">Competition mathematics</div></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);"><strong>41.7</strong></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">1.04</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">40.42</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">34.17</td></tr><tr><td style="padding:6px 4px 6px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle"><div style="font-size:12.5px;font-weight:600;line-height:1.2;color:inherit">AIME 2026</div><div style="margin-top:2px;font-size:10px;opacity:0.65">Competition mathematics</div></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);"><strong>48.5</strong></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">0.21</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">40.42</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">38.75</td></tr><tr><td style="padding:6px 4px 6px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle"><div style="font-size:12.5px;font-weight:600;line-height:1.2;color:inherit">HMMT Feb 2026</div><div style="margin-top:2px;font-size:10px;opacity:0.65">Competition mathematics</div></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);"><strong>25.8</strong></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">0.57</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">23.3</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">22.73</td></tr><tr><td colspan="5" style="padding:6px 10px;font-weight:600;font-size:12.5px;color:#2450D6;border-bottom:1px solid rgba(36,80,214,0.25);background:rgba(36,80,214,0.12)">Scientific Reasoning</td></tr><tr><td style="padding:6px 4px 6px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle"><div style="font-size:12.5px;font-weight:600;line-height:1.2;color:inherit">GPQA Diamond</div><div style="margin-top:2px;font-size:10px;opacity:0.65">Graduate-level science QA</div></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);"><strong>27.3</strong></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">11.9</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">26.26</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">54.89</td></tr><tr><td colspan="5" style="padding:6px 10px;font-weight:600;font-size:12.5px;color:#2450D6;border-bottom:1px solid rgba(36,80,214,0.25);background:rgba(36,80,214,0.12)">Coding</td></tr><tr><td style="padding:6px 4px 6px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle"><div style="font-size:12.5px;font-weight:600;line-height:1.2;color:inherit">HumanEval+</div><div style="margin-top:2px;font-size:10px;opacity:0.65">Code generation</div></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);"><strong>79.9</strong></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">16.5</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">65.2</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">75.6</td></tr><tr><td style="padding:6px 4px 6px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle"><div style="font-size:12.5px;font-weight:600;line-height:1.2;color:inherit">MBPP+</div><div style="margin-top:2px;font-size:10px;opacity:0.65">Code generation</div></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);"><strong>68</strong></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">35.4</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">60.6</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">67.7</td></tr><tr><td style="padding:6px 4px 6px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle"><div style="font-size:12.5px;font-weight:600;line-height:1.2;color:inherit">LiveCodeBench v6</div><div style="margin-top:2px;font-size:10px;opacity:0.65">Competitive coding</div></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);"><strong>37.41</strong></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">6.64</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">33.52</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">29.79</td></tr><tr><td colspan="5" style="padding:6px 10px;font-weight:600;font-size:12.5px;color:#2450D6;border-bottom:1px solid rgba(36,80,214,0.25);background:rgba(36,80,214,0.12)">Agents</td></tr><tr><td style="padding:6px 4px 6px 10px;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle"><div style="font-size:12.5px;font-weight:600;line-height:1.2;color:inherit">BFCL v4</div><div style="margin-top:2px;font-size:10px;opacity:0.65">Function calling</div></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;background:rgba(36,80,214,0.08);"><strong>28</strong></td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">25.3</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">25.15</td><td style="padding:6px 2px;text-align:center;border-bottom:1px solid rgba(128,128,128,0.15);vertical-align:middle;font-size:12.5px;line-height:1.2;">43.6</td></tr></tbody></table></div>
50
  <!-- TABLE:END -->
51
 
52
+ Scores in %. Bold highlights K2-Horizon-0.9B; Qwen3.5-2B is included as a larger reference model. Protocol and provenance details are in the [Technical Appendix](APPENDIX.md#evaluation).
53
 
54
  ## Quickstart
55
 
 
62
 
63
  ```shell
64
  vllm serve IFM/K2-Horizon-0.9B \
 
 
 
65
  --trust-remote-code \
66
  --dtype bfloat16 \
67
  --max-model-len 131072 \
68
+ --hf-overrides '{"rope_parameters":{"rope_type":"yarn","factor":16.0,"original_max_position_embeddings":8192}}' \
69
+ --gpu-memory-utilization 0.85 \
70
+ --tensor-parallel-size 1 \
71
  --reasoning-parser k2_horizon \
72
  --enable-auto-tool-choice \
73
  --tool-call-parser k2_horizon
 
76
  SGLang, from a source checkout that includes [sgl-project/sglang#37654](https://github.com/sgl-project/sglang/pull/37654). This is the recipe validated in the [SGLang K2 Horizon cookbook](https://docs.sglang.io/cookbook/autoregressive/IFM/K2-Horizon):
77
 
78
  ```shell
79
+ sglang serve \
80
  --model-path IFM/K2-Horizon-0.9B \
81
+ --revision 9b9ec1f7e17f62ed218df542687a144116219d84 \
82
  --tp 1 \
83
  --dtype bfloat16 \
84
  --attention-backend fa3 \
85
  --reasoning-parser k2_horizon \
86
+ --host 0.0.0.0 \
87
+ --port 30000
88
  ```
89
 
90
  ### API Usage
 
132
  ## Best Practices
133
 
134
  1. **Reasoning effort: always `high`.** All reported results use high reasoning effort. Pass `{"chat_template_kwargs": {"reasoning_effort": "high"}}` on every request; `medium` and `low` trade accuracy for speed and are not recommended for evaluation.
135
+ 2. **Sampling parameters.** `temperature=0.6`, `top_p=0.95`.
136
  3. **Output length.** Allow at least 32,768 output tokens so reasoning is never cut off. Truncated reasoning is a failed response, not a shorter one.
137
  4. **Serving.** Use the validated SGLang recipe above: BF16, TP=1, FlashAttention-3. Full recipes for every K2-Horizon size, with measured H200 latency and throughput, are in the [SGLang cookbook](https://docs.sglang.io/cookbook/autoregressive/IFM/K2-Horizon).
138
  5. **Parsers.** Enable the `k2_horizon` reasoning parser for chat, and add the `k2_horizon` tool-call parser for agent use. Leave both off for plain completion-style generation.
139
+ 6. **Revisions.** `main` is the MOPD release checkpoint; `mid1_75k` and `mid2_47k` preserve the context-extension stages.
140
 
141
  ## Citation
142
 
 
149
  }
150
  ```
151
 
152
+