Add Artificial Analysis evaluations for kimi-k2-thinking
Browse filesThis commit adds structured evaluation results to the model card. The results are formatted using the model-index specification and will be displayed in the model card's evaluation widget.
README.md
CHANGED
|
@@ -1,281 +1,332 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: other
|
| 3 |
-
license_name: modified-mit
|
| 4 |
-
library_name: transformers
|
| 5 |
-
|
| 6 |
-
|
| 7 |
-
|
| 8 |
-
|
| 9 |
-
|
| 10 |
-
|
| 11 |
-
|
| 12 |
-
|
| 13 |
-
|
| 14 |
-
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
| 25 |
-
|
| 26 |
-
|
| 27 |
-
|
| 28 |
-
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
|
| 34 |
-
|
| 35 |
-
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
| 49 |
-
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
|
| 65 |
-
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
| 79 |
-
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
|
| 87 |
-
|
| 88 |
-
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
|
| 92 |
-
|
| 93 |
-
|
| 94 |
-
|
| 95 |
-
|
| 96 |
-
|
| 97 |
-
|
| 98 |
-
|
|
| 99 |
-
|
| 100 |
-
**
|
| 101 |
-
|
|
| 102 |
-
|
| 103 |
-
| **
|
| 104 |
-
| **
|
| 105 |
-
| **
|
| 106 |
-
| **
|
| 107 |
-
| **
|
| 108 |
-
| **
|
| 109 |
-
| **
|
| 110 |
-
|
| 111 |
-
|
| 112 |
-
|
| 113 |
-
|
| 114 |
-
|
| 115 |
-
|
| 116 |
-
|
| 117 |
-
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
|
| 136 |
-
|
| 137 |
-
|
| 138 |
-
|
| 139 |
-
|
| 140 |
-
|
| 141 |
-
|
| 142 |
-
|
| 143 |
-
|
| 144 |
-
|
| 145 |
-
|
| 146 |
-
|
| 147 |
-
|
| 148 |
-
|
| 149 |
-
|
| 150 |
-
|
| 151 |
-
|
| 152 |
-
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
*
|
| 156 |
-
*
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
|
| 160 |
-
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
|
| 164 |
-
|
| 165 |
-
|
| 166 |
-
|
| 167 |
-
|
| 168 |
-
|
| 169 |
-
|
| 170 |
-
|
| 171 |
-
|
| 172 |
-
|
| 173 |
-
|
| 174 |
-
|
| 175 |
-
|
| 176 |
-
|
| 177 |
-
|
| 178 |
-
|
| 179 |
-
|
| 180 |
-
|
| 181 |
-
|
| 182 |
-
|
| 183 |
-
|
| 184 |
-
|
| 185 |
-
|
| 186 |
-
|
| 187 |
-
|
| 188 |
-
|
| 189 |
-
|
| 190 |
-
|
| 191 |
-
|
| 192 |
-
|
| 193 |
-
|
| 194 |
-
|
| 195 |
-
|
| 196 |
-
|
| 197 |
-
|
| 198 |
-
|
| 199 |
-
|
| 200 |
-
|
| 201 |
-
|
| 202 |
-
|
| 203 |
-
|
| 204 |
-
|
| 205 |
-
|
| 206 |
-
|
| 207 |
-
|
| 208 |
-
|
| 209 |
-
|
| 210 |
-
|
| 211 |
-
|
| 212 |
-
|
| 213 |
-
|
| 214 |
-
|
| 215 |
-
|
| 216 |
-
|
| 217 |
-
|
| 218 |
-
|
| 219 |
-
|
| 220 |
-
|
| 221 |
-
|
| 222 |
-
|
| 223 |
-
|
| 224 |
-
|
| 225 |
-
|
| 226 |
-
|
| 227 |
-
|
| 228 |
-
|
| 229 |
-
|
| 230 |
-
|
| 231 |
-
|
| 232 |
-
|
| 233 |
-
|
| 234 |
-
|
| 235 |
-
|
| 236 |
-
|
| 237 |
-
|
| 238 |
-
|
| 239 |
-
|
| 240 |
-
|
| 241 |
-
|
| 242 |
-
|
| 243 |
-
|
| 244 |
-
|
| 245 |
-
|
| 246 |
-
|
| 247 |
-
|
| 248 |
-
|
| 249 |
-
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
|
| 253 |
-
|
| 254 |
-
|
| 255 |
-
|
| 256 |
-
|
| 257 |
-
|
| 258 |
-
|
| 259 |
-
|
| 260 |
-
|
| 261 |
-
|
| 262 |
-
|
| 263 |
-
|
| 264 |
-
|
| 265 |
-
|
| 266 |
-
|
| 267 |
-
|
| 268 |
-
|
| 269 |
-
|
| 270 |
-
|
| 271 |
-
|
| 272 |
-
|
| 273 |
-
|
| 274 |
-
|
| 275 |
-
|
| 276 |
-
|
| 277 |
-
|
| 278 |
-
|
| 279 |
-
|
| 280 |
-
|
| 281 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: other
|
| 3 |
+
license_name: modified-mit
|
| 4 |
+
library_name: transformers
|
| 5 |
+
model-index:
|
| 6 |
+
- name: Kimi-K2-Thinking
|
| 7 |
+
results:
|
| 8 |
+
- task:
|
| 9 |
+
type: evaluation
|
| 10 |
+
dataset:
|
| 11 |
+
name: Artificial Analysis Benchmarks
|
| 12 |
+
type: artificial_analysis
|
| 13 |
+
metrics:
|
| 14 |
+
- name: Artificial Analysis Intelligence Index
|
| 15 |
+
type: artificial_analysis_intelligence_index
|
| 16 |
+
value: 67
|
| 17 |
+
- name: Artificial Analysis Coding Index
|
| 18 |
+
type: artificial_analysis_coding_index
|
| 19 |
+
value: 52.2
|
| 20 |
+
- name: Artificial Analysis Math Index
|
| 21 |
+
type: artificial_analysis_math_index
|
| 22 |
+
value: 94.7
|
| 23 |
+
- name: Mmlu Pro
|
| 24 |
+
type: mmlu_pro
|
| 25 |
+
value: 0.848
|
| 26 |
+
- name: Gpqa
|
| 27 |
+
type: gpqa
|
| 28 |
+
value: 0.838
|
| 29 |
+
- name: Hle
|
| 30 |
+
type: hle
|
| 31 |
+
value: 0.223
|
| 32 |
+
- name: Livecodebench
|
| 33 |
+
type: livecodebench
|
| 34 |
+
value: 0.853
|
| 35 |
+
- name: Scicode
|
| 36 |
+
type: scicode
|
| 37 |
+
value: 0.424
|
| 38 |
+
- name: Aime 25
|
| 39 |
+
type: aime_25
|
| 40 |
+
value: 0.947
|
| 41 |
+
- name: Ifbench
|
| 42 |
+
type: ifbench
|
| 43 |
+
value: 0.681
|
| 44 |
+
- name: Lcr
|
| 45 |
+
type: lcr
|
| 46 |
+
value: 0.663
|
| 47 |
+
- name: Terminalbench Hard
|
| 48 |
+
type: terminalbench_hard
|
| 49 |
+
value: 0.291
|
| 50 |
+
- name: Tau2
|
| 51 |
+
type: tau2
|
| 52 |
+
value: 0.93
|
| 53 |
+
source:
|
| 54 |
+
name: Artificial Analysis API
|
| 55 |
+
url: https://artificialanalysis.ai
|
| 56 |
+
---
|
| 57 |
+
<div align="center">
|
| 58 |
+
<picture>
|
| 59 |
+
<img src="figures/kimi-logo.png" width="30%" alt="Kimi K2: Open Agentic Intellignece">
|
| 60 |
+
</picture>
|
| 61 |
+
</div>
|
| 62 |
+
<hr>
|
| 63 |
+
|
| 64 |
+
<div align="center" style="line-height:1">
|
| 65 |
+
<a href="https://www.kimi.com" target="_blank"><img alt="Chat" src="https://img.shields.io/badge/🤖%20Chat-Kimi%20K2-ff6b6b?color=1783ff&logoColor=white"/></a>
|
| 66 |
+
<a href="https://www.moonshot.ai" target="_blank"><img alt="Homepage" src="https://img.shields.io/badge/Homepage-Moonshot%20AI-white?logo=Kimi&logoColor=white"/></a>
|
| 67 |
+
</div>
|
| 68 |
+
|
| 69 |
+
<div align="center" style="line-height: 1;">
|
| 70 |
+
<a href="https://huggingface.co/moonshotai" target="_blank"><img alt="Hugging Face" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Moonshot%20AI-ffc107?color=ffc107&logoColor=white"/></a>
|
| 71 |
+
<a href="https://twitter.com/kimi_moonshot" target="_blank"><img alt="Twitter Follow" src="https://img.shields.io/badge/Twitter-Kimi.ai-white?logo=x&logoColor=white"/></a>
|
| 72 |
+
<a href="https://discord.gg/TYU2fdJykW" target="_blank"><img alt="Discord" src="https://img.shields.io/badge/Discord-Kimi.ai-white?logo=discord&logoColor=white"/></a>
|
| 73 |
+
</div>
|
| 74 |
+
<div align="center" style="line-height: 1;">
|
| 75 |
+
<a href="https://huggingface.co/moonshotai/Kimi-K2-Thinking/blob/main/LICENSE"><img alt="License" src="https://img.shields.io/badge/License-Modified_MIT-f5de53?&color=f5de53"/></a>
|
| 76 |
+
</div>
|
| 77 |
+
|
| 78 |
+
<p align="center">
|
| 79 |
+
<b>📰 <a href="https://moonshotai.github.io/Kimi-K2/thinking.html">Tech Blog</a></b>
|
| 80 |
+
</p>
|
| 81 |
+
|
| 82 |
+
|
| 83 |
+
## 1. Model Introduction
|
| 84 |
+
|
| 85 |
+
Kimi K2 Thinking is the latest, most capable version of open-source thinking model. Starting with Kimi K2, we built it as a thinking agent that reasons step-by-step while dynamically invoking tools. It sets a new state-of-the-art on Humanity's Last Exam (HLE), BrowseComp, and other benchmarks by dramatically scaling multi-step reasoning depth and maintaining stable tool-use across 200–300 sequential calls. At the same time, K2 Thinking is a native INT4 quantization model with 256k context window, achieving lossless reductions in inference latency and GPU memory usage.
|
| 86 |
+
|
| 87 |
+
### Key Features
|
| 88 |
+
- **Deep Thinking & Tool Orchestration**: End-to-end trained to interleave chain-of-thought reasoning with function calls, enabling autonomous research, coding, and writing workflows that last hundreds of steps without drift.
|
| 89 |
+
- **Native INT4 Quantization**: Quantization-Aware Training (QAT) is employed in post-training stage to achieve lossless 2x speed-up in low-latency mode.
|
| 90 |
+
- **Stable Long-Horizon Agency**: Maintains coherent goal-directed behavior across up to 200–300 consecutive tool invocations, surpassing prior models that degrade after 30–50 steps.
|
| 91 |
+
|
| 92 |
+
|
| 93 |
+
## 2. Model Summary
|
| 94 |
+
|
| 95 |
+
<div align="center">
|
| 96 |
+
|
| 97 |
+
|
| 98 |
+
| | |
|
| 99 |
+
|:---:|:---:|
|
| 100 |
+
| **Architecture** | Mixture-of-Experts (MoE) |
|
| 101 |
+
| **Total Parameters** | 1T |
|
| 102 |
+
| **Activated Parameters** | 32B |
|
| 103 |
+
| **Number of Layers** (Dense layer included) | 61 |
|
| 104 |
+
| **Number of Dense Layers** | 1 |
|
| 105 |
+
| **Attention Hidden Dimension** | 7168 |
|
| 106 |
+
| **MoE Hidden Dimension** (per Expert) | 2048 |
|
| 107 |
+
| **Number of Attention Heads** | 64 |
|
| 108 |
+
| **Number of Experts** | 384 |
|
| 109 |
+
| **Selected Experts per Token** | 8 |
|
| 110 |
+
| **Number of Shared Experts** | 1 |
|
| 111 |
+
| **Vocabulary Size** | 160K |
|
| 112 |
+
| **Context Length** | 256K |
|
| 113 |
+
| **Attention Mechanism** | MLA |
|
| 114 |
+
| **Activation Function** | SwiGLU |
|
| 115 |
+
</div>
|
| 116 |
+
|
| 117 |
+
## 3. Evaluation Results
|
| 118 |
+
|
| 119 |
+
**Reasoning Tasks**
|
| 120 |
+
| Benchmark | Setting | K2 Thinking | GPT-5<br> (High) | Claude Sonnet 4.5<br> (Thinking) | K2 0905 | DeepSeek-V3.2 | Grok-4 |
|
| 121 |
+
|:----------:|:--------:|:------------:|:------:|:----------------------------:|:--------:|:--------------:|:-------:|
|
| 122 |
+
| **HLE (Text-only)** | no tools | 23.9 | 26.3 | 19.8* | 7.9 | 19.8 | 25.4 |
|
| 123 |
+
| | w/ tools | 44.9 | 41.7* | 32.0* | 21.7 | 20.3* | 41.0 |
|
| 124 |
+
| | heavy | 51.0 | 42.0 | - | - | - | 50.7 |
|
| 125 |
+
| **AIME25** | no tools | 94.5 | 94.6 | 87.0 | 51.0 | 89.3 | 91.7 |
|
| 126 |
+
| | w/ python | 99.1 | 99.6 | 100.0 | 75.2 | 58.1* | 98.8 |
|
| 127 |
+
| | heavy | 100.0 | 100.0 | - | - | - | 100.0 |
|
| 128 |
+
| **HMMT25** | no tools | 89.4 | 93.3 | 74.6* | 38.8 | 83.6 | 90.0 |
|
| 129 |
+
| | w/ python | 95.1 | 96.7 | 88.8* | 70.4 | 49.5* | 93.9 |
|
| 130 |
+
| | heavy | 97.5 | 100.0 | - | - | - | 96.7 |
|
| 131 |
+
| **IMO-AnswerBench** | no tools | 78.6 | 76.0* | 65.9* | 45.8 | 76.0* | 73.1 |
|
| 132 |
+
| **GPQA** | no tools | 84.5 | 85.7 | 83.4 | 74.2 | 79.9 | 87.5 |
|
| 133 |
+
|
| 134 |
+
**General Tasks**
|
| 135 |
+
| Benchmark | Setting | K2 Thinking | GPT-5<br> (High) | Claude Sonnet 4.5<br> (Thinking) | K2 0905 | DeepSeek-V3.2 |
|
| 136 |
+
|:----------:|:--------:|:------------:|:------:|:----------------------------:|:--------:|:--------------:|
|
| 137 |
+
| **MMLU-Pro** | no tools | 84.6 | 87.1 | 87.5 | 81.9 | 85.0 |
|
| 138 |
+
| **MMLU-Redux** | no tools | 94.4 | 95.3 | 95.6 | 92.7 | 93.7 |
|
| 139 |
+
| **Longform Writing** | no tools | 73.8 | 71.4 | 79.8 | 62.8 | 72.5 |
|
| 140 |
+
| **HealthBench** | no tools | 58.0 | 67.2 | 44.2 | 43.8 | 46.9 |
|
| 141 |
+
|
| 142 |
+
**Agentic Search Tasks**
|
| 143 |
+
| Benchmark | Setting | K2 Thinking | GPT-5<br> (High) | Claude Sonnet 4.5<br> (Thinking) | K2 0905 | DeepSeek-V3.2 |
|
| 144 |
+
|:----------:|:--------:|:------------:|:------:|:----------------------------:|:--------:|:--------------:|
|
| 145 |
+
| **BrowseComp** | w/ tools | 60.2 | 54.9 | 24.1 | 7.4 | 40.1 |
|
| 146 |
+
| **BrowseComp-ZH** | w/ tools | 62.3 | 63.0* | 42.4* | 22.2 | 47.9 |
|
| 147 |
+
| **Seal-0** | w/ tools | 56.3 | 51.4* | 53.4* | 25.2 | 38.5* |
|
| 148 |
+
| **FinSearchComp-T3** | w/ tools | 47.4 | 48.5* | 44.0* | 10.4 | 27.0* |
|
| 149 |
+
| **Frames** | w/ tools | 87.0 | 86.0* | 85.0* | 58.1 | 80.2* |
|
| 150 |
+
|
| 151 |
+
**Coding Tasks**
|
| 152 |
+
| Benchmark | Setting | K2 Thinking | GPT-5<br> (High) | Claude Sonnet 4.5<br> (Thinking) | K2 0905 | DeepSeek-V3.2 |
|
| 153 |
+
|:----------:|:--------:|:------------:|:------:|:----------------------------:|:--------:|:--------------:|
|
| 154 |
+
| **SWE-bench Verified** | w/ tools | 71.3 | 74.9 | 77.2 | 69.2 | 67.8 |
|
| 155 |
+
| **SWE-bench Multilingual** | w/ tools | 61.1 | 55.3* | 68.0 | 55.9 | 57.9 |
|
| 156 |
+
| **Multi-SWE-bench** | w/ tools | 41.9 | 39.3* | 44.3 | 33.5 | 30.6 |
|
| 157 |
+
| **SciCode** | no tools | 44.8 | 42.9 | 44.7 | 30.7 | 37.7 |
|
| 158 |
+
| **LiveCodeBenchV6** | no tools | 83.1 | 87.0* | 64.0* | 56.1* | 74.1 |
|
| 159 |
+
| **OJ-Bench (cpp)** | no tools | 48.7 | 56.2* | 30.4* | 25.5* | 38.2* |
|
| 160 |
+
| **Terminal-Bench** | w/ simulated tools (JSON) | 47.1 | 43.8 | 51.0 | 44.5 | 37.7 |
|
| 161 |
+
<details>
|
| 162 |
+
<summary><b>Footnotes</b></summary>
|
| 163 |
+
|
| 164 |
+
1. To ensure a fast, lightweight experience, we selectively employ a subset of tools and reduce the number of tool call steps under the chat mode on kimi.com. As a result, chatting on kimi.com may not reproduce our benchmark scores. Our agentic mode will be updated soon to reflect the full capabilities of K2 Thinking.
|
| 165 |
+
|
| 166 |
+
2. **Testing Details**:
|
| 167 |
+
2.1. All benchmarks were evaluated at temperature = 1.0 and 256 k context length for K2 Thinking, except for SciCode, for which we followed the official temperature setting of 0.0.
|
| 168 |
+
2.2. HLE (no tools), AIME25, HMMT25, and GPQA were capped at a 96k thinking-token budget, while IMO-Answer Bench, LiveCodeBench and OJ-Bench were capped at a 128k thinking-token budget. Longform Writing was capped at a 32k completion-token budget.
|
| 169 |
+
2.3. For AIME and HMMT (no tools), we report the average of 32 runs (avg@32). For AIME and HMMT (with Python), we report the average of 16 runs (avg@16). For IMO-AnswerBench, we report the average of 8 runs (avg@8).
|
| 170 |
+
|
| 171 |
+
3. **Baselines**:
|
| 172 |
+
3.1 GPT-5, Claude-4.5-sonnet, Grok-4 results and DeepSeek-V3.2 results are quoted from the [GPT-5 post](https://openai.com/index/introducing-gpt-5/), [GPT-5 for Developers post](https://openai.com/index/introducing-gpt-5-for-developers/), [GPT-5 system card](https://openai.com/index/gpt-5-system-card/), [claude-sonnet-4-5 post](https://www.anthropic.com/news/claude-sonnet-4-5), [grok-4 post](https://x.ai/news/grok-4), [deepseek-v3.2 post](https://api-docs.deepseek.com/news/news250929), the [public Terminal-Bench leaderboard](https://www.tbench.ai/leaderboard) (Terminus-2), the [public Vals AI leaderboard](https://vals.ai/) and [artificialanalysis](https://artificialanalysis.ai/). Benchmarks for which no available public scores were re-tested under the same conditions used for k2 thinking and are marked with an asterisk(*). For the GPT-5 test, we set the reasoning effort to high.
|
| 173 |
+
3.2 The GPT-5 and Grok-4 on the HLE full set with tools are 35.2 and 38.6 from the official posts. In our internal evaluation on the HLE text-only subset, GPT-5 scores 41.7 and Grok-4 scores 38.6 (Grok-4’s launch cited 41.0 on the text-only subset). For GPT-5's HLE text-only w/o tool, we use score from <a href="https://scale.com/leaderboard/humanitys_last_exam_text_only" target="_blank">Scale.ai</a>. The official GPT5 HLE full set w/o tool is 24.8.
|
| 174 |
+
3.3 For <a href="https://aclanthology.org/2025.emnlp-main.1794.pdf" target="_blank">IMO-AnswerBench</a>: GPT-5 scored 65.6 in the benchmark paper. We re-evaluated GPT-5 with official API and obtained a score of 76.
|
| 175 |
+
|
| 176 |
+
4. **For HLE (w/ tools) and the agentic-search benchmarks**:
|
| 177 |
+
4.1. K2 Thinking was equipped with search, code-interpreter, and web-browsing tools.
|
| 178 |
+
4.2. BrowseComp-ZH, Seal-0 and FinSearchComp-T3 were run 4 times independently and the average is reported (avg@4).
|
| 179 |
+
4.3. The evaluation used o3-mini as judge, configured identically to the official HLE setting; judge prompts were taken verbatim from the official repository.
|
| 180 |
+
4.4. On HLE, the maximum step limit was 120, with a 48 k-token reasoning budget per step; on agentic-search tasks, the limit was 300 steps with a 24 k-token reasoning budget per step.
|
| 181 |
+
4.5. When tool execution results cause the accumulated input to exceed the model's context limit (256k), we employ a simple context management strategy that hides all previous tool outputs.
|
| 182 |
+
4.6. The web access to Hugging Face may lead to data leakage in certain benchmark tests, such as HLE. K2 Thinking can achieve a score of 51.3 on HLE without blocking Hugging Face. To ensure a fair and rigorous comparison, we blocked access to Hugging Face during testing.
|
| 183 |
+
|
| 184 |
+
5. **For Coding Tasks**:
|
| 185 |
+
5.1. Terminal-Bench scores were obtained with the default agent framework (Terminus-2) and the provided JSON parser.
|
| 186 |
+
5.2. For other coding tasks, the result was produced with our in-house evaluation harness. The harness is derived from SWE-agent, but we clamp the context windows of the Bash and Edit tools and rewrite the system prompt to match the task semantics.
|
| 187 |
+
5.3. All reported scores of coding tasks are averaged over 5 independent runs.
|
| 188 |
+
|
| 189 |
+
6. **Heavy Mode**: K2 Thinking Heavy Mode employs an efficient parallel strategy: it first rolls out eight trajectories simultaneously, then reflectively aggregates all outputs to generate the final result. Heavy mode for GPT-5 denotes the official GPT-5 Pro score.
|
| 190 |
+
</details>
|
| 191 |
+
|
| 192 |
+
## 4. Native INT4 Quantization
|
| 193 |
+
|
| 194 |
+
Low-bit quantization is an effective way to reduce inference latency and GPU memory usage on large-scale inference servers. However, thinking models use excessive decoding lengths, and thus quantization often results in substantial performance drops.
|
| 195 |
+
|
| 196 |
+
To overcome this challenge, we adopt Quantization-Aware Training (QAT) during the post-training phase, applying INT4 weight-only quantization to the MoE components. It allows K2 Thinking to support native INT4 inference with a roughly 2x generation speed improvement while achieving state-of-the-art performance. All benchmark results are reported under INT4 precision.
|
| 197 |
+
|
| 198 |
+
The checkpoints are saved in compressed-tensors format, supported by most of mainstream inference engine. If you need the checkpoints in higher precision such as FP8 or BF16, you can refer to [official repo of compressed-tensors](https://github.com/vllm-project/compressed-tensors) to unpack the int4 weights and convert to any higher precision.
|
| 199 |
+
|
| 200 |
+
## 5. Deployment
|
| 201 |
+
> [!Note]
|
| 202 |
+
> You can access K2 Thinking's API on https://platform.moonshot.ai , we provide OpenAI/Anthropic-compatible API for you.
|
| 203 |
+
|
| 204 |
+
Currently, Kimi-K2-Thinking is recommended to run on the following inference engines:
|
| 205 |
+
|
| 206 |
+
* vLLM
|
| 207 |
+
* SGLang
|
| 208 |
+
* KTransformers
|
| 209 |
+
|
| 210 |
+
Deployment examples can be found in the [Model Deployment Guide](docs/deploy_guidance.md).
|
| 211 |
+
|
| 212 |
+
---
|
| 213 |
+
|
| 214 |
+
## 6. Model Usage
|
| 215 |
+
|
| 216 |
+
### Chat Completion
|
| 217 |
+
|
| 218 |
+
Once the local inference service is up, you can interact with it through the chat endpoint:
|
| 219 |
+
|
| 220 |
+
```python
|
| 221 |
+
def simple_chat(client: openai.OpenAI, model_name: str):
|
| 222 |
+
messages = [
|
| 223 |
+
{"role": "system", "content": "You are Kimi, an AI assistant created by Moonshot AI."},
|
| 224 |
+
{"role": "user", "content": [{"type": "text", "text": "which one is bigger, 9.11 or 9.9? think carefully."}]},
|
| 225 |
+
]
|
| 226 |
+
response = client.chat.completions.create(
|
| 227 |
+
model=model_name,
|
| 228 |
+
messages=messages,
|
| 229 |
+
stream=False,
|
| 230 |
+
temperature=1.0,
|
| 231 |
+
max_tokens=4096
|
| 232 |
+
)
|
| 233 |
+
print(f"k2 answer: {response.choices[0].message.content}")
|
| 234 |
+
print("=====below is reasoning content======")
|
| 235 |
+
print(f"reasoning content: {response.choices[0].message.reasoning_content}")
|
| 236 |
+
```
|
| 237 |
+
|
| 238 |
+
> [!NOTE]
|
| 239 |
+
> The recommended temperature for Kimi-K2-Thinking is `temperature = 1.0`.
|
| 240 |
+
> If no special instructions are required, the system prompt above is a good default.
|
| 241 |
+
|
| 242 |
+
---
|
| 243 |
+
|
| 244 |
+
### Tool Calling
|
| 245 |
+
|
| 246 |
+
Kimi-K2-Thinking has the same tool calling settings as Kimi-K2-Instruct.
|
| 247 |
+
|
| 248 |
+
To enable them, you need to pass the list of available tools in each request, then the model will autonomously decide when and how to invoke them.
|
| 249 |
+
|
| 250 |
+
The following example demonstrates calling a weather tool end-to-end:
|
| 251 |
+
|
| 252 |
+
```python
|
| 253 |
+
# Your tool implementation
|
| 254 |
+
def get_weather(city: str) -> dict:
|
| 255 |
+
return {"weather": "Sunny"}
|
| 256 |
+
# Tool schema definition
|
| 257 |
+
tools = [{
|
| 258 |
+
"type": "function",
|
| 259 |
+
"function": {
|
| 260 |
+
"name": "get_weather",
|
| 261 |
+
"description": "Retrieve current weather information. Call this when the user asks about the weather.",
|
| 262 |
+
"parameters": {
|
| 263 |
+
"type": "object",
|
| 264 |
+
"required": ["city"],
|
| 265 |
+
"properties": {
|
| 266 |
+
"city": {
|
| 267 |
+
"type": "string",
|
| 268 |
+
"description": "Name of the city"
|
| 269 |
+
}
|
| 270 |
+
}
|
| 271 |
+
}
|
| 272 |
+
}
|
| 273 |
+
}]
|
| 274 |
+
# Map tool names to their implementations
|
| 275 |
+
tool_map = {
|
| 276 |
+
"get_weather": get_weather
|
| 277 |
+
}
|
| 278 |
+
def tool_call_with_client(client: OpenAI, model_name: str):
|
| 279 |
+
messages = [
|
| 280 |
+
{"role": "system", "content": "You are Kimi, an AI assistant created by Moonshot AI."},
|
| 281 |
+
{"role": "user", "content": "What's the weather like in Beijing today? Use the tool to check."}
|
| 282 |
+
]
|
| 283 |
+
finish_reason = None
|
| 284 |
+
while finish_reason is None or finish_reason == "tool_calls":
|
| 285 |
+
completion = client.chat.completions.create(
|
| 286 |
+
model=model_name,
|
| 287 |
+
messages=messages,
|
| 288 |
+
temperature=1.0,
|
| 289 |
+
tools=tools, # tool list defined above
|
| 290 |
+
tool_choice="auto"
|
| 291 |
+
)
|
| 292 |
+
choice = completion.choices[0]
|
| 293 |
+
finish_reason = choice.finish_reason
|
| 294 |
+
if finish_reason == "tool_calls":
|
| 295 |
+
messages.append(choice.message)
|
| 296 |
+
for tool_call in choice.message.tool_calls:
|
| 297 |
+
tool_call_name = tool_call.function.name
|
| 298 |
+
tool_call_arguments = json.loads(tool_call.function.arguments)
|
| 299 |
+
tool_function = tool_map[tool_call_name]
|
| 300 |
+
tool_result = tool_function(**tool_call_arguments)
|
| 301 |
+
print("tool_result:", tool_result)
|
| 302 |
+
messages.append({
|
| 303 |
+
"role": "tool",
|
| 304 |
+
"tool_call_id": tool_call.id,
|
| 305 |
+
"name": tool_call_name,
|
| 306 |
+
"content": json.dumps(tool_result)
|
| 307 |
+
})
|
| 308 |
+
print("-" * 100)
|
| 309 |
+
print(choice.message.content)
|
| 310 |
+
```
|
| 311 |
+
|
| 312 |
+
The `tool_call_with_client` function implements the pipeline from user query to tool execution.
|
| 313 |
+
This pipeline requires the inference engine to support Kimi-K2’s native tool-parsing logic.
|
| 314 |
+
For more information, see the [Tool Calling Guide](docs/tool_call_guidance.md).
|
| 315 |
+
|
| 316 |
+
---
|
| 317 |
+
|
| 318 |
+
## 7. License
|
| 319 |
+
|
| 320 |
+
Both the code repository and the model weights are released under the [Modified MIT License](LICENSE).
|
| 321 |
+
|
| 322 |
+
---
|
| 323 |
+
|
| 324 |
+
## 8. Third Party Notices
|
| 325 |
+
|
| 326 |
+
See [THIRD PARTY NOTICES](THIRD_PARTY_NOTICES.md)
|
| 327 |
+
|
| 328 |
+
---
|
| 329 |
+
|
| 330 |
+
## 9. Contact Us
|
| 331 |
+
|
| 332 |
+
If you have any questions, please reach out at [[email protected]](mailto:[email protected]).
|