Hugging Face's logo

moonshotai
/
Kimi-K2-Instruct

Text Generation
Transformers
Safetensors
kimi_k2
conversational
custom_code
fp8
Model card Files Community
6
lsw825 commited on
Commit
8a112f5
·
verified ·
1 Parent(s): 22bf139

Upload folder using huggingface_hub

Browse files
Files changed (7) hide show
  1. .gitattributes +1 -0
  2. LICENSE +27 -0
  3. README.md +665 -3
  4. docs/deploy_guidance.md +196 -0
  5. docs/tool_call_guidance.md +241 -0
  6. figures/Base-Evaluation.png +3 -0
  7. figures/kimi-logo.png +0 -0
.gitattributes CHANGED
@@ -34,3 +34,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
  model.safetensors.index.json filter=lfs diff=lfs merge=lfs -text
 
 
34
  *.zst filter=lfs diff=lfs merge=lfs -text
35
  *tfevents* filter=lfs diff=lfs merge=lfs -text
36
  model.safetensors.index.json filter=lfs diff=lfs merge=lfs -text
37
+ figures/Base-Evaluation.png filter=lfs diff=lfs merge=lfs -text
LICENSE ADDED
@@ -0,0 +1,27 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ Modified MIT License
2
+
3
+ Copyright (c) 2025 Moonshot AI
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the “Software”), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
22
+
23
+ Our only modification part is that, if the Software (or any derivative works
24
+ thereof) is used for any of your commercial products or services that have
25
+ more than 100 million monthly active users, or more than 20 million US dollars
26
+ (or equivalent in other currencies) in monthly revenue, you shall prominently
27
+ display "Kimi K2" on the user interface of such product or service.
README.md CHANGED
@@ -1,3 +1,665 @@
1
- ---
2
- license: mit
3
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ <!-- ---
2
+ library_name: transformers
3
+ --- -->
4
+ <!-- markdownlint-disable first-line-h1 -->
5
+ <!-- markdownlint-disable html -->
6
+ <!-- markdownlint-disable no-duplicate-header -->
7
+
8
+
9
+ <div align="center">
10
+ <picture>
11
+ <img src="figures/kimi-logo.png" width="30%" alt="Kimi K2: Open Agentic Intellignece">
12
+ </picture>
13
+ </div>
14
+ <hr>
15
+ <div align="center" style="line-height:1">
16
+ <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>
17
+ <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>
18
+ </div>
19
+
20
+ <div align="center" style="line-height: 1;">
21
+ <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>
22
+ <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>
23
+ <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>
24
+ </div>
25
+
26
+ <div align="center" style="line-height: 1;">
27
+ <a href="https://github.com/moonshotai/Kimi-K2/blob/main/LICENSE"><img alt="License" src="https://img.shields.io/badge/License-Modified_MIT-f5de53?&color=f5de53"/></a>
28
+ </div>
29
+
30
+ <p align="center">
31
+ <b>📰&nbsp;&nbsp;<a href="https://moonshotai.github.io/Kimi-K2/">Tech Blog</a></b> &nbsp;&nbsp;&nbsp; | &nbsp;&nbsp;&nbsp; <b>📄&nbsp;&nbsp;Paper Link (comming soon)</b>
32
+ </p>
33
+
34
+ ## 0. Reminder: Remove this after you squash the commit history before release.
35
+
36
+ ## 1. Model Introduction
37
+
38
+ Kimi K2 is a state-of-the-art mixture-of-experts (MoE) language model with 32 billion activated parameters and 1 trillion total parameters. Trained with the Muon optimizer, Kimi K2 achieves exceptional performance across frontier knowledge, reasoning, and coding tasks while being meticulously optimized for agentic capabilities.
39
+
40
+ ### Key Features
41
+ - Large-Scale Training: Pre-trained a 1T parameter MoE model on 15.5T tokens with zero training instability.
42
+ - MuonClip Optimizer: We apply the Muon optimizer to an unprecedented scale, and develop novel optimization techniques to resolve instabilities while scaling up.
43
+ - Agentic Intelligence: Specifically designed for tool use, reasoning, and autonomous problem-solving.
44
+
45
+ ### Model Variants
46
+ - **Kimi-K2-Base**: The foundation model, a strong start for researchers and builders who want full control for fine-tuning and custom solutions.
47
+ - **Kimi-K2-Instruct**: The post-trained model best for drop-in, general-purpose chat and agentic experiences. It is a reflex-grade model without long thinking.
48
+
49
+
50
+ <p align="center">
51
+ TODO this is a banner
52
+ <img width="80%" src="figures/logo.svg">
53
+ </p>
54
+
55
+ ## 2. Model Summary
56
+
57
+ <div align="center">
58
+
59
+
60
+ | | |
61
+ |:---:|:---:|
62
+ | **Architecture** | Mixture-of-Experts (MoE) |
63
+ | **Total Parameters** | 1T |
64
+ | **Activated Parameters** | 32B |
65
+ | **Number of Layers** (Dense layer included) | 61 |
66
+ | **Number of Dense Layers** | 1 |
67
+ | **Attention Hidden Dimension** | 7168 |
68
+ | **MoE Hidden Dimension** (per Expert) | 2048 |
69
+ | **Number of Attention Heads** | 64 |
70
+ | **Number of Experts** | 384 |
71
+ | **Selected Experts per Token** | 8 |
72
+ | **Number of Shared Experts** | 1 |
73
+ | **Vocabulary Size** | 160K |
74
+ | **Context Length** | 128K |
75
+ | **Attention Mechanism** | MLA |
76
+ | **Activation Function** | SwiGLU |
77
+ </div>
78
+
79
+ ## 3. Evaluation Results
80
+
81
+ #### Instruction model evaluation results
82
+
83
+ <div align="center">
84
+ <table>
85
+ <thead>
86
+ <tr>
87
+ <th align="center">Benchmark</th>
88
+ <th align="center">Metric</th>
89
+ <th align="center">Kimi K2 Instruct</th>
90
+ <th align="center">DeepSeek-V3-0324</th>
91
+ <th align="center">Qwen3-235B-A22B <br><sup>(non-thinking)</sup></th>
92
+ <th align="center">Claude Sonnet 4 <br><sup>(w/o extended thinking)</sup></th>
93
+ <th align="center">Claude Opus 4 <br><sup>(w/o extended thinking)</sup></th>
94
+ <th align="center">GPT-4.1</th>
95
+ <th align="center">Gemini 2.5 Flash <br> Preview (05-20)</th>
96
+ </tr>
97
+ </thead>
98
+ <tbody>
99
+ <tr>
100
+ <td align="center" colspan=9><strong>Coding Tasks</strong></td>
101
+ </tr>
102
+ <tr>
103
+ <td align="center">LiveCodeBench v6<br><sup>(Aug 24 - May 25)</sup></td>
104
+ <td align="center">Pass@1</td>
105
+ <td align="center"><strong>53.7</strong></td>
106
+ <td align="center">46.9</td>
107
+ <td align="center">37.0</td>
108
+ <td align="center">48.5</td>
109
+ <td align="center">47.4</t6
110
+ <td align="center">44.7</td>
111
+ <td align="center">44.7</td>
112
+ </tr>
113
+ <tr>
114
+ <td align="center">OJBench</td>
115
+ <td align="center">Pass@1</td>
116
+ <td align="center"><strong>27.1</strong></td>
117
+ <td align="center">24.0</td>
118
+ <td align="center">11.3</td>
119
+ <td align="center">15.3</td>
120
+ <td align="center">19.6</td>
121
+ <td align="center">19.5</td>
122
+ <td align="center">19.5</td>
123
+ </tr>
124
+ <tr>
125
+ <td align="center">MultiPL-E</td>
126
+ <td align="center">Pass@1</td>
127
+ <td align="center"><ins><strong>86.7</strong></ins></td>
128
+ <td align="center">83.1</td>
129
+ <td align="center">78.2</td>
130
+ <td align="center">88.6</td>
131
+ <td align="center"><strong>89.6</strong></td>
132
+ <td align="center">86.7</td>
133
+ <td align="center">85.6</td>
134
+ </tr>
135
+ <tr>
136
+ <td align="center">SWE-bench Verified <br/><sup>(Agentless Coding)</sup></td>
137
+ <td align="center">Single Patch</td>
138
+ <td align="center"><ins><strong>51.8</strong></ins></td>
139
+ <td align="center">36.6</td>
140
+ <td align="center">39.4</td>
141
+ <td align="center">50.2</td>
142
+ <td align="center"><strong>53.0</strong></td>
143
+ <td align="center">40.8</td>
144
+ <td align="center">32.6</td>
145
+ </tr>
146
+
147
+ <tr>
148
+ <td align="center" rowspan="2">SWE-bench Verified <br/> <sup>(Agentic Coding)</sup></td>
149
+ <td align="center">Single Attempt (Acc)</td>
150
+ <td align="center"><ins><strong>65.8</strong></ins></td>
151
+ <td align="center">38.8</td>
152
+ <td align="center">34.4</td>
153
+ <td align="center"><strong>72.7</strong></td>
154
+ <td align="center">72.5<sup>*</sup></td>
155
+ <td align="center">54.6</td>
156
+ <td align="center">—</td>
157
+ </tr>
158
+ <tr>
159
+ <!--<td align="center">(Agentic Coding)</td>-->
160
+ <td align="center">Multiple Attempts (Acc)</td>
161
+ <td align="center"><ins><strong>71.6</strong></ins></td>
162
+ <td align="center">—</td>
163
+ <td align="center">—</td>
164
+ <td align="center"><strong>80.2</strong></td>
165
+ <td align="center">79.4<sup>*</sup></td>
166
+ <td align="center">—</td>
167
+ <td align="center">—</td>
168
+ </tr>
169
+
170
+ <tr>
171
+ <td align="center" rowspan="2">SWE-bench Multilingual<br /> <sup>(Agentic Coding)</sup></td>
172
+ <td align="center">Single Attempt (Acc)</td>
173
+ <td align="center"><ins><strong>47.3</strong> </ins></td>
174
+ <td align="center">25.8</td>
175
+ <td align="center">20.9</td>
176
+ <td align="center"><strong>51.0</strong></td>
177
+ <td align="center">—</td>
178
+ <td align="center">31.5</td>
179
+ <td align="center">—</td>
180
+ </tr>
181
+ <tr>
182
+ <!--<td align="center">(Agentic Coding)</td>-->
183
+ <td align="center">Inhouse Framework (Acc)</td>
184
+ <td align="center"><ins><strong>30.0</strong> </ins></td>
185
+ <td align="center">—</td>
186
+ <td align="center">—</td>
187
+ <td align="center">35.5</td>
188
+ <td align="center"><strong>43.2</strong></td>
189
+ <td align="center">8.30</td>
190
+ <td align="center">—</td>
191
+ </tr>
192
+ <tr>
193
+ <td align="center">TerminalBench</td>
194
+ <td align="center">Acc</td>
195
+ <td align="center"><ins><strong>25.0</strong> </ins></td>
196
+ <td align="center">16.3</td>
197
+ <td align="center">6.60</td>
198
+ <td align="center">—</td>
199
+ <td align="center">—</td>
200
+ <td align="center"><strong>30.3</strong></td>
201
+ <td align="center">16.8</td>
202
+ </tr>
203
+ <tr>
204
+ <td align="center">Aider-Polyglot</td>
205
+ <td align="center">Acc</td>
206
+ <td align="center">60.0</td>
207
+ <td align="center">55.1</td>
208
+ <td align="center"><ins><strong>61.8</strong></ins></td>
209
+ <td align="center">56.4</td>
210
+ <td align="center"><strong>70.7</strong></td>
211
+ <td align="center">52.4</td>
212
+ <td align="center">44.0</td>
213
+ </tr>
214
+ <tr>
215
+ <td align="center" colspan=9><strong>Tool Use Tasks</strong></td>
216
+ </tr>
217
+ <tr>
218
+ <td align="center">Tau2 retail</td>
219
+ <td align="center">Avg@4</td>
220
+ <td align="center"><ins><strong>70.6</strong></ins></td>
221
+ <td align="center">69.1</td>
222
+ <td align="center">57.0</td>
223
+ <td align="center">75.0</td>
224
+ <td align="center"><strong>81.8</strong></td>
225
+ <td align="center">74.8</td>
226
+ <td align="center">64.3</td>
227
+ </tr>
228
+ <tr>
229
+ <td align="center">Tau2 airline</td>
230
+ <td align="center">Avg@4</td>
231
+ <td align="center"><ins><strong>56.5</strong></ins></td>
232
+ <td align="center">39.0</td>
233
+ <td align="center">26.5</td>
234
+ <td align="center">55.5</td>
235
+ <td align="center"><strong>60.0</strong></td>
236
+ <td align="center">54.5</td>
237
+ <td align="center">42.5</td>
238
+ </tr>
239
+ <tr>
240
+ <td align="center">Tau2 telecom</td>
241
+ <td align="center">Avg@4</td>
242
+ <td align="center"><strong>65.8</strong></td>
243
+ <td align="center">32.5</td>
244
+ <td align="center">22.1</td>
245
+ <td align="center">45.2</td>
246
+ <td align="center">57.0</td>
247
+ <td align="center">38.6</td>
248
+ <td align="center">16.9</td>
249
+ </tr>
250
+ <tr>
251
+ <td align="center">AceBench</td>
252
+ <td align="center">Acc</td>
253
+ <td align="center"><ins><strong>76.5</strong></ins></td>
254
+ <td align="center">72.7</td>
255
+ <td align="center">70.5</td>
256
+ <td align="center">76.2</td>
257
+ <td align="center">75.6</td>
258
+ <td align="center"><strong>80.1</strong></td>
259
+ <td align="center">74.5</td>
260
+ </tr>
261
+ <tr>
262
+ <td align="center" colspan=9><strong>Math &amp; STEM Tasks</strong></td>
263
+ </tr>
264
+ <tr>
265
+ <td align="center">AIME 2024</td>
266
+ <td align="center">Avg@64</td>
267
+ <td align="center"><strong>69.6</strong></td>
268
+ <td align="center">59.4<sup>*</sup></td>
269
+ <td align="center">40.1<sup>*</sup></td>
270
+ <td align="center">43.4</td>
271
+ <td align="center">48.2</td>
272
+ <td align="center">46.5</td>
273
+ <td align="center">61.3</td>
274
+ </tr>
275
+ <tr>
276
+ <td align="center">AIME 2025</td>
277
+ <td align="center">Avg@64</td>
278
+ <td align="center"><strong>49.5</strong></td>
279
+ <td align="center">46.7</td>
280
+ <td align="center">24.7<sup>*</sup></td>
281
+ <td align="center">33.1<sup>*</sup></td>
282
+ <td align="center">33.9<sup>*</sup></td>
283
+ <td align="center">37.0</td>
284
+ <td align="center">46.6</td>
285
+ </tr>
286
+ <tr>
287
+ <td align="center">MATH-500</td>
288
+ <td align="center">Acc</td>
289
+ <td align="center"><strong>97.4</strong></td>
290
+ <td align="center">94.0<sup>*</sup></td>
291
+ <td align="center">91.2<sup>*</sup></td>
292
+ <td align="center">94.0</td>
293
+ <td align="center">94.4</td>
294
+ <td align="center">92.4/td>
295
+ <td align="center">95.4</td>
296
+ </tr>
297
+ <tr>
298
+ <td align="center">HMMT 2025</td>
299
+ <td align="center">Avg@32</td>
300
+ <td align="center"><strong>38.8</strong></td>
301
+ <td align="center">27.5</td>
302
+ <td align="center">11.9</td>
303
+ <td align="center">15.9</td>
304
+ <td align="center">15.8</td>
305
+ <td align="center">19.4</td>
306
+ <td align="center">34.7</td>
307
+ </tr>
308
+ <tr>
309
+ <td align="center">CNMO 2024</td>
310
+ <td align="center">Avg@16</td>
311
+ <td align="center">74.3</td>
312
+ <td align="center"><ins><strong>74.7</strong></ins></td>
313
+ <td align="center">48.6</td>
314
+ <td align="center">60.4</td>
315
+ <td align="center">57.6</td>
316
+ <td align="center">56.6</td>
317
+ <td align="center"><strong>75.0</strong></td>
318
+ </tr>
319
+ <tr>
320
+ <td align="center">PolyMath-en</td>
321
+ <td align="center">Avg@4</td>
322
+ <td align="center"><strong>65.1</strong></td>
323
+ <td align="center">59.5</td>
324
+ <td align="center">51.9</td>
325
+ <td align="center">52.8</td>
326
+ <td align="center">49.8</td>
327
+ <td align="center">54.0</td>
328
+ <td align="center">49.9</td>
329
+ </tr>
330
+
331
+ <tr>
332
+ <td align="center">ZebraLogic</td>
333
+ <td align="center">Acc</td>
334
+ <td align="center"><strong>89.0</strong></td>
335
+ <td align="center">84.0</td>
336
+ <td align="center">37.7</td>
337
+ <td align="center">73.7</td>
338
+ <td align="center">59.3</td>
339
+ <td align="center">58.5</td>
340
+ <td align="center">57.9</td>
341
+ </tr>
342
+
343
+ <tr>
344
+ <td align="center">AutoLogi</td>
345
+ <td align="center">Acc</td>
346
+ <td align="center"><ins><strong>89.5</strong></ins></td>
347
+ <td align="center">88.9</td>
348
+ <td align="center">83.3</td>
349
+ <td align="center"><strong>89.8</strong></td>
350
+ <td align="center">86.1</td>
351
+ <td align="center">88.2</td>
352
+ <td align="center">84.1</td>
353
+ </tr>
354
+
355
+ <tr>
356
+ <td align="center">GPQA-Diamond</td>
357
+ <td align="center">Avg@8</td>
358
+ <td align="center"><strong>75.1</strong></td>
359
+ <td align="center">68.4<sup>*</sup></td>
360
+ <td align="center">62.9<sup>*</sup></td>
361
+ <td align="center">70.0<sup>*</sup></td>
362
+ <td align="center">74.9<sup>*</sup></td>
363
+ <td align="center">66.3</td>
364
+ <td align="center">68.2</td>
365
+ </tr>
366
+
367
+ <tr>
368
+ <td align="center">SuperGPQA</td>
369
+ <td align="center">Acc</td>
370
+ <td align="center"><strong>57.2</strong></td>
371
+ <td align="center">53.7</td>
372
+ <td align="center">50.2</td>
373
+ <td align="center">55.7</td>
374
+ <td align="center">56.5</td>
375
+ <td align="center">50.8</td>
376
+ <td align="center">49.6</td>
377
+ </tr>
378
+
379
+ <tr>
380
+ <td align="center">Humanity’s Last</td>
381
+ <td align="center">(Text Only)</td>
382
+ <td align="center">4.7</td>
383
+ <td align="center">5.2</td>
384
+ <td align="center"><ins><strong>5.7</strong></ins></td>
385
+ <td align="center">5.8</td>
386
+ <td align="center"><strong>7.1</strong></td>
387
+ <td align="center">3.7</td>
388
+ <td align="center">5.6</td>
389
+ </tr>
390
+
391
+ <tr>
392
+ <td align="center" colspan=9><strong>General Tasks</strong></td>
393
+ </tr>
394
+
395
+ <tr>
396
+ <td align="center">MMLU</td>
397
+ <td align="center">EM</td>
398
+ <td align="center"><ins><strong>89.5</strong></ins></td>
399
+ <td align="center">89.4</td>
400
+ <td align="center">87.0</td>
401
+ <td align="center">91.5</td>
402
+ <td align="center"><strong>92.9</strong></td>
403
+ <td align="center">90.4</td>
404
+ <td align="center">90.1</td>
405
+ </tr>
406
+
407
+ <tr>
408
+ <td align="center">MMLU-Redux</td>
409
+ <td align="center">EM</td>
410
+ <td align="center"><ins><strong>92.7</strong></ins></td>
411
+ <td align="center">90.5</td>
412
+ <td align="center">89.2</td>
413
+ <td align="center">93.6</td>
414
+ <td align="center"><strong>94.2</strong></td>
415
+ <td align="center">92.4</td>
416
+ <td align="center">90.6</td>
417
+ </tr>
418
+
419
+ <tr>
420
+ <td align="center">MMLU-Pro</td>
421
+ <td align="center">EM</td>
422
+ <td align="center">81.1</td>
423
+ <td align="center"><ins><strong>81.2</strong></ins><sup>*</sup></td>
424
+ <td align="center">77.3</td>
425
+ <td align="center">83.7</td>
426
+ <td align="center"><strong>86.6</strong></td>
427
+ <td align="center">81.8</td>
428
+ <td align="center">79.4</td>
429
+ </tr>
430
+
431
+ <tr>
432
+ <td align="center">IFEval</td>
433
+ <td align="center">Prompt Strict</td>
434
+ <td align="center"><strong>89.8</strong></td>
435
+ <td align="center">81.1</td>
436
+ <td align="center">83.2<sup>*</sup></td>
437
+ <td align="center">87.6</td>
438
+ <td align="center">87.4</td>
439
+ <td align="center">88.0</td>
440
+ <td align="center">84.3</td>
441
+ </tr>
442
+
443
+ <tr>
444
+ <td align="center">Multi-Challenge</td>
445
+ <td align="center">Acc</td>
446
+ <td align="center"><strong>54.1</strong></td>
447
+ <td align="center">31.4</td>
448
+ <td align="center">34.0</td>
449
+ <td align="center">46.8</td>
450
+ <td align="center">49.0</td>
451
+ <td align="center">36.4</td>
452
+ <td align="center">39.5</td>
453
+ </tr>
454
+
455
+ <tr>
456
+ <td align="center">SimpleQA</td>
457
+ <td align="center">Correct</td>
458
+ <td align="center"><ins><strong>31.0</strong></ins></td>
459
+ <td align="center">27.7</td>
460
+ <td align="center">13.2</td>
461
+ <td align="center">15.9</td>
462
+ <td align="center">22.8</td>
463
+ <td align="center"><strong>42.3</strong></td>
464
+ <td align="center">23.3</td>
465
+ </tr>
466
+
467
+ <tr>
468
+ <td align="center">Livebench</td>
469
+ <td align="center">Pass@1</td>
470
+ <td align="center"><strong>76.4</strong></td>
471
+ <td align="center">72.4</td>
472
+ <td align="center">67.6</td>
473
+ <td align="center">74.8</td>
474
+ <td align="center">74.6</td>
475
+ <td align="center">69.8</td>
476
+ <td align="center">67.8</td>
477
+ </tr>
478
+ </tbody>
479
+ </table>
480
+ </div>
481
+ <sup>
482
+ • Bold denotes global SOTA, and underlined denotes open-source SOTA.
483
+ </sup><br/><sup>
484
+ • Data points marked with * are taken directly from the model's tech report or blog.
485
+ </sup><br/><sup>
486
+ • All metrics, except for SWE-bench Verified (Agentless), are evaluated with an 8k output token length. SWE-bench Verified (Agentless) is limited to a 16k output token length.
487
+ </sup><br/><sup>
488
+ • Kimi K2 achieves 65.8% pass@1 on the SWE-bench Verified tests with bash/editor tools (single-attempt patches, no test-time compute). It also achieves a 47.3% pass@1 on the SWE-bench Multilingual tests under the same conditions. Additionally, we report results on SWE-bench Verified tests (71.6%) that leverage parallel test-time compute by sampling multiple sequences and selecting the single best via an internal scoring model.
489
+ </sup><br/><sup>
490
+ • To ensure the stability of the evaluation, we employed avg@k on the AIME, HMMT, CNMO, PolyMath-en, GPQA-Diamond, EvalPlus, Tau2.
491
+ </sup><br/><sup>
492
+ • Some data points have been omitted due to prohibitively expensive evaluation costs.
493
+ </sup>
494
+
495
+ ---
496
+
497
+ #### Base model evaluation results
498
+
499
+ <div align="center">
500
+
501
+ | Benchmark | Metric | Shot | Kimi K2 Base | Deepseek-V3-Base | Qwen2.5-72B | Llama 4 Maverick |
502
+ |:-------------------:|:----------:|:---------:|:--------------:|:------------------:|:-------------:|:------------------:|
503
+ | **General Tasks** | | | | | | |
504
+ | MMLU | EM | 5-shot | **87.79** | 87.1 | 86.08 | 84.87 |
505
+ | MMLU-pro | EM | 5-shot | **69.17** | 60.59 | 62.8 | 63.47 |
506
+ | MMLU-redux-2.0 | EM | 5-shot | **90.17** | 89.53 | 87.77 | 88.18 |
507
+ | SimpleQA | Correct | 5-shot | **35.25** | 26.49 | 10.31 | 23.74 |
508
+ | TriviaQA | EM | 5-shot | **85.09** | 84.11 | 76.03 | 79.25 |
509
+ | GPQA-Diamond | Avg@8 | 5-shot | 48.11 | **50.51** | 40.78 | 49.43 |
510
+ | SuperGPQA | EM | 5-shot | **44.67** | 39.2 | 34.23 | 38.84 |
511
+ | **Code Tasks** | | | | | | |
512
+ | LiveCodeBench v6 | Pass@1 | 1-shot | **26.29** | 22.86 | 21.14 | 25.14 |
513
+ | EvalPlus | Pass@1 | - | **80.33** | 65.61 | 66.04 | 65.48 |
514
+ | **Mathematics Tasks** | | | | | | |
515
+ | MATH | EM | 4-shot | **70.22** | 60.06 | 60.96 | 63.02 |
516
+ | GSM8k | EM | 8-shot | **92.12** | 91.66 | 90.37 | 86.35 |
517
+ | **Chinese Tasks** | | | | | | |
518
+ | C-Eval | EM | 5-shot | **92.5** | 90.04 | 90.86 | 80.91 |
519
+ | CSimpleQA | Correct | 5-shot | **77.57** | 72.13 | 50.53 | 53.47 |
520
+
521
+ </div>
522
+ <sup>
523
+ • We only evaluate open-source pretrained models in this work. We report results for Qwen2.5-72B because the base checkpoint for Qwen3-235B-A22B was not open-sourced at the time of our study.
524
+ </sup><br/><sup>
525
+ • All models are evaluated using the same evaluation protocol.
526
+
527
+ </sup>
528
+
529
+
530
+ ## 4. Deployment
531
+ > [!Note]
532
+ > You can access Kimi K2's API on https://platform.moonshot.ai , we provide OpenAI/Anthropic-compatible API for you.
533
+ >
534
+ > The Anthropic-compatible API maps temperature by `real_temperature = request_temperature * 0.6` for better compatible with existing applications.
535
+
536
+ Our model checkpoints are stored in the block-fp8 format, you can find it on [Huggingface](https://huggingface.co/moonshotai/Kimi-K2-Instruct).
537
+
538
+ Currently, Kimi-K2 is recommended to run on the following inference engines:
539
+
540
+ * vLLM
541
+ * SGLang
542
+ * KTransformers
543
+ * TensorRT-LLM
544
+
545
+ Deployment examples for vLLM and SGLang can be found in the [Model Deployment Guide](docs/deploy_guidance.md).
546
+
547
+ ---
548
+
549
+ ## 5. Model Usage
550
+
551
+ ### Chat Completion
552
+
553
+ Once the local inference service is up, you can interact with it through the chat endpoint:
554
+
555
+ ```python
556
+ def simple_chat(client: OpenAI, model_name: str):
557
+ messages = [
558
+ {"role": "system", "content": "You are Kimi, an AI assistant created by Moonshot AI."},
559
+ {"role": "user", "content": [{"type": "text", "text": "Please give a brief self-introduction."}]},
560
+ ]
561
+ response = client.chat.completions.create(
562
+ model=model_name,
563
+ messages=messages,
564
+ stream=False,
565
+ temperature=0.6,
566
+ max_tokens=256
567
+ )
568
+ print(response.choices[0].message.content)
569
+ ```
570
+
571
+ > [!NOTE]
572
+ > The recommended temperature for Kimi-K2-Instruct is `temperature = 0.6`.
573
+ > If no special instructions are required, the system prompt above is a good default.
574
+
575
+ ---
576
+
577
+ ### Tool Calling
578
+
579
+ Kimi-K2-Instruct has strong tool-calling capabilities.
580
+ 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.
581
+
582
+ The following example demonstrates calling a weather tool end-to-end:
583
+
584
+ ```python
585
+ # Your tool implementation
586
+ def get_weather(city: str) -> dict:
587
+ return {"weather": "Sunny"}
588
+
589
+ # Tool schema definition
590
+ tools = [{
591
+ "type": "function",
592
+ "function": {
593
+ "name": "get_weather",
594
+ "description": "Retrieve current weather information. Call this when the user asks about the weather.",
595
+ "parameters": {
596
+ "type": "object",
597
+ "required": ["city"],
598
+ "properties": {
599
+ "city": {
600
+ "type": "string",
601
+ "description": "Name of the city"
602
+ }
603
+ }
604
+ }
605
+ }
606
+ }]
607
+
608
+ # Map tool names to their implementations
609
+ tool_map = {
610
+ "get_weather": get_weather
611
+ }
612
+
613
+ def tool_call_with_client(client: OpenAI, model_name: str):
614
+ messages = [
615
+ {"role": "system", "content": "You are Kimi, an AI assistant created by Moonshot AI."},
616
+ {"role": "user", "content": "What's the weather like in Beijing today? Use the tool to check."}
617
+ ]
618
+ finish_reason = None
619
+ while finish_reason is None or finish_reason == "tool_calls":
620
+ completion = client.chat.completions.create(
621
+ model=model_name,
622
+ messages=messages,
623
+ temperature=0.6,
624
+ tools=tools, # tool list defined above
625
+ tool_choice="auto"
626
+ )
627
+ choice = completion.choices[0]
628
+ finish_reason = choice.finish_reason
629
+ if finish_reason == "tool_calls":
630
+ messages.append(choice.message)
631
+ for tool_call in choice.message.tool_calls:
632
+ tool_call_name = tool_call.function.name
633
+ tool_call_arguments = json.loads(tool_call.function.arguments)
634
+ tool_function = tool_map[tool_call_name]
635
+ tool_result = tool_function(**tool_call_arguments)
636
+ print("tool_result:", tool_result)
637
+
638
+ messages.append({
639
+ "role": "tool",
640
+ "tool_call_id": tool_call.id,
641
+ "name": tool_call_name,
642
+ "content": json.dumps(tool_result)
643
+ })
644
+ print("-" * 100)
645
+ print(choice.message.content)
646
+ ```
647
+
648
+ The `tool_call_with_client` function implements the pipeline from user query to tool execution.
649
+ This pipeline requires the inference engine to support Kimi-K2’s native tool-parsing logic.
650
+ For streaming output and manual tool-parsing, see the [Tool Calling Guide](docs/tool_call_guidance.md).
651
+
652
+ ---
653
+
654
+ ## 6. License
655
+
656
+ Both the code repository and the model weights are released under the [Modified MIT License](LICENSE).
657
+
658
+ In short, it is MIT License for most people, but you need to give credit to "Kimi K2" by displaying it prominently in your product, if you have more than 100 million monthly active users or annual revenue exceeding 20 million USD.
659
+
660
+
661
+ ---
662
+
663
+ ## 7. Contact Us
664
+
665
+ If you have any questions, please reach out at [[email protected]](mailto:[email protected]).
docs/deploy_guidance.md ADDED
@@ -0,0 +1,196 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Kimi-K2 Deployment Guide
2
+
3
+ > [!Note]
4
+ > This guide only provides some examples of deployment commands for Kimi-K2, which may not be the optimal configuration. Since inference engines are still being updated frequenty, please continue to follow the guidance from their homepage if you want to achieve better inference performance.
5
+
6
+
7
+ ## vLLM Deployment
8
+
9
+ The smallest deployment unit for Kimi-K2 FP8 weights with 128k seqlen on mainstream H200 or H20 platform is a cluster with 16 GPUs with either Tensor Parallel (TP) or "data parallel + expert parallel" (DP+EP).
10
+ Running parameters for this environment are provided below. You may scale up to more nodes and increase expert-parallelism to enlarge the inference batch size and overall throughput.
11
+
12
+ ### Tensor Parallelism
13
+
14
+ When the parallelism degree ≤ 16, you can run inference with pure Tensor Parallelism. A sample launch command is:
15
+
16
+ ``` bash
17
+ # start ray on node 0 and node 1
18
+
19
+ # node 0:
20
+ vllm serve $MODEL_PATH \
21
+ --port 8000 \
22
+ --served-model-name kimi-k2 \
23
+ --trust-remote-code \
24
+ --tensor-parallel-size 16 \
25
+ --enable-auto-tool-choice \
26
+ --tool-call-parser kimi_k2
27
+ ```
28
+
29
+ **Key parameter notes:**
30
+ - `--tensor-parallel-size 16`: If using more than 16 GPUs, combine with pipeline-parallelism.
31
+ - `--enable-auto-tool-choice`: Required when enabling tool usage.
32
+ - `--tool-call-parser kimi_k2`: Required when enabling tool usage.
33
+
34
+ ### Data Parallelism + Expert Parallelism
35
+
36
+ You can install libraries like DeepEP and DeepGEMM as needed. Then run the command (example on H200):
37
+
38
+ ``` bash
39
+ # node 0
40
+ vllm serve $MODEL_PATH --port 8000 --served-model-name kimi-k2 --trust-remote-code --data-parallel-size 16 --data-parallel-size-local 8 --data-parallel-address $MASTER_IP --data-parallel-rpc-port $PORT --enable-expert-parallel --max-num-batched-tokens 8192 --max-num-seqs 256 --gpu-memory-utilization 0.85 --enable-auto-tool-choice --tool-call-parser kimi_k2
41
+
42
+ # node 1
43
+ vllm serve $MODEL_PATH --headless --data-parallel-start-rank 8 --port 8000 --served-model-name kimi-k2 --trust-remote-code --data-parallel-size 16 --data-parallel-size-local 8 --data-parallel-address $MASTER_IP --data-parallel-rpc-port $PORT --enable-expert-parallel --max-num-batched-tokens 8192 --max-num-seqs 256 --gpu-memory-utilization 0.85 --enable-auto-tool-choice --tool-call-parser kimi_k2
44
+ ```
45
+
46
+ ## SGLang Deployment
47
+
48
+ Similarly, we can use TP or DP+EP in SGLang for Deployment, here are the examples.
49
+
50
+
51
+ ### Tensor Parallelism
52
+
53
+ Here is the simple example code to run TP16 with two nodes on H200:
54
+
55
+ ``` bash
56
+ # Node 0
57
+ python -m sglang.launch_server --model-path $MODEL_PATH --tp 16 --dist-init-addr $MASTER_IP:50000 --nnodes 2 --node-rank 0 --trust-remote-code --tool-call-parser kimi_k2
58
+
59
+ # Node 1
60
+ python -m sglang.launch_server --model-path $MODEL_PATH --tp 16 --dist-init-addr $MASTER_IP:50000 --nnodes 2 --node-rank 1 --trust-remote-code --tool-call-parser kimi_k2
61
+ ```
62
+
63
+ **Key parameter notes:**
64
+ - `--tool-call-parser kimi_k2`: Required when enabling tool usage.
65
+
66
+ ### Data Parallelism + Expert Parallelism
67
+
68
+ Here is an example for large scale Prefill-Decode Disagreegation (4P12D H200) with DP+EP in SGLang:
69
+
70
+ ``` bash
71
+ # for prefill node
72
+ MC_TE_METRIC=true SGLANG_DISAGGREGATION_HEARTBEAT_INTERVAL=10000000 SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=100000 SGLANG_DISAGGREGATION_WAITING_TIMEOUT=100000 PYTHONUNBUFFERED=1 \
73
+ python -m sglang.launch_server --model-path $MODEL_PATH \
74
+ --trust-remote-code --disaggregation-mode prefill --dist-init-addr $PREFILL_NODE0$:5757 --tp-size 32 --dp-size 32 --enable-dp-attention --host $LOCAL_IP --decode-log-interval 1 --disable-radix-cache --enable-deepep-moe --moe-dense-tp-size 1 --enable-dp-lm-head --disable-shared-experts-fusion --watchdog-timeout 1000000 --enable-two-batch-overlap --disaggregation-ib-device $IB_DEVICE --chunked-prefill-size 131072 --mem-fraction-static 0.85 --deepep-mode normal --ep-dispatch-algorithm dynamic --eplb-algorithm deepseek --max-running-requests 1024 --nnodes 4 --node-rank $RANK --tool-call-parser kimi_k2
75
+
76
+
77
+ # for decode node
78
+ SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK=480 MC_TE_METRIC=true SGLANG_DISAGGREGATION_HEARTBEAT_INTERVAL=10000000 SGLANG_DISAGGREGATION_BOOTSTRAP_TIMEOUT=100000 SGLANG_DISAGGREGATION_WAITING_TIMEOUT=100000 PYTHONUNBUFFERED=1 \
79
+ python -m sglang.launch_server --model-path $MODEL_PATH --trust-remote-code --disaggregation-mode decode --dist-init-addr $DECODE_NODE0:5757 --tp-size 96 --dp-size 96 --enable-dp-attention --host $LOCAL_IP --decode-log-interval 1 --context-length 2176 --disable-radix-cache --enable-deepep-moe --moe-dense-tp-size 1 --enable-dp-lm-head --disable-shared-experts-fusion --watchdog-timeout 1000000 --enable-two-batch-overlap --disaggregation-ib-device $IB_DEVICE --deepep-mode low_latency --mem-fraction-static 0.8 --cuda-graph-bs 480 --max-running-requests 46080 --ep-num-redundant-experts 96 --nnodes 12 --node-rank $RANK --tool-call-parser kimi_k2
80
+
81
+ # pdlb
82
+ PYTHONUNBUFFERED=1 python -m sglang.srt.disaggregation.launch_lb --prefill http://${PREFILL_NODE0}:30000 --decode http://${DECODE_NODE0}:30000
83
+ ```
84
+
85
+ ## KTransformers Deployment
86
+
87
+ Please copy all configuration files (i.e., everything except the .safetensors files) into the GGUF checkpoint folder at /path/to/K2. Then run:
88
+ ``` bash
89
+ python ktransformers/server/main.py --model_path /path/to/K2 --gguf_path /path/to/K2 --cache_lens 30000
90
+ ```
91
+
92
+ To enable AMX optimization, run:
93
+
94
+ ``` bash
95
+ python ktransformers/server/main.py --model_path /path/to/K2 --gguf_path /path/to/K2 --cache_lens 30000 --optimize_config_path ktransformers/optimize/optimize_rules/DeepSeek-V3-Chat-fp8-linear-ggml-experts-serve-amx.yaml
96
+ ```
97
+
98
+ ## TensoRT-LLM Deployment
99
+ ### Prerequisite
100
+ Please refer to [this guide](https://nvidia.github.io/TensorRT-LLM/installation/build-from-source-linux.html) to build TensorRT-LLM v1.0.0-rc2 from source and start a TRT-LLM docker container.
101
+
102
+ install blobfile by:
103
+ ```bash
104
+ pip install blobfile
105
+ ```
106
+ ### Multi-node Serving
107
+ TensorRT-LLM supports multi-node inference. You can use mpirun to launch Kimi-K2 with multi-node jobs. We will use two nodes for this example.
108
+
109
+ #### mpirun
110
+ mpirun requires each node to have passwordless ssh access to the other node. We need to setup the environment inside the docker container. Run the container with host network and mount the current directory as well as model directory to the container.
111
+
112
+ ```bash
113
+ # use host network
114
+ IMAGE=<YOUR_IMAGE>
115
+ NAME=test_2node_docker
116
+ # host1
117
+ docker run -it --name ${NAME}_host1 --ipc=host --gpus=all --network host --privileged --ulimit memlock=-1 --ulimit stack=67108864 -v ${PWD}:/workspace -v <YOUR_MODEL_DIR>:/models/DeepSeek-V3 -w /workspace ${IMAGE}
118
+ # host2
119
+ docker run -it --name ${NAME}_host2 --ipc=host --gpus=all --network host --privileged --ulimit memlock=-1 --ulimit stack=67108864 -v ${PWD}:/workspace -v <YOUR_MODEL_DIR>:/models/DeepSeek-V3 -w /workspace ${IMAGE}
120
+ ```
121
+
122
+ Set up ssh inside the container
123
+
124
+ ```bash
125
+ apt-get update && apt-get install -y openssh-server
126
+
127
+ # modify /etc/ssh/sshd_config
128
+ PermitRootLogin yes
129
+ PubkeyAuthentication yes
130
+ # modify /etc/ssh/sshd_config, change default port 22 to another unused port
131
+ port 2233
132
+
133
+ # modify /etc/ssh
134
+ ```
135
+
136
+ Generate ssh key on host1 and copy to host2, vice versa.
137
+
138
+ ```bash
139
+ # on host1
140
+ ssh-keygen -t ed25519 -f ~/.ssh/id_ed25519
141
+ ssh-copy-id -i ~/.ssh/id_ed25519.pub root@<HOST2>
142
+ # on host2
143
+ ssh-keygen -t ed25519 -f ~/.ssh/id_ed25519
144
+ ssh-copy-id -i ~/.ssh/id_ed25519.pub root@<HOST1>
145
+
146
+ # restart ssh service on host1 and host2
147
+ service ssh restart # or
148
+ /etc/init.d/ssh restart # or
149
+ systemctl restart ssh
150
+ ```
151
+
152
+ Generate additional config for trtllm serve.
153
+ ```bash
154
+ cat >/path/to/TensorRT-LLM/extra-llm-api-config.yml <<EOF
155
+ cuda_graph_config:
156
+ padding_enabled: true
157
+ batch_sizes:
158
+ - 1
159
+ - 2
160
+ - 4
161
+ - 8
162
+ - 16
163
+ - 32
164
+ - 64
165
+ - 128
166
+ print_iter_log: true
167
+ enable_attention_dp: true
168
+ EOF
169
+ ```
170
+
171
+
172
+ After the preparations,you can run the trtllm-serve on two nodes using mpirun:
173
+
174
+ ```bash
175
+ mpirun -np 16 \
176
+ -H <HOST1>:8,<HOST2>:8 \
177
+ -mca plm_rsh_args "-p 2233" \
178
+ --allow-run-as-root \
179
+ trtllm-llmapi-launch trtllm-serve serve \
180
+ --backend pytorch \
181
+ --tp_size 16 \
182
+ --ep_size 8 \
183
+ --kv_cache_free_gpu_memory_fraction 0.95 \
184
+ --trust_remote_code \
185
+ --max_batch_size 128 \
186
+ --max_num_tokens 4096 \
187
+ --extra_llm_api_options /path/to/TensorRT-LLM/extra-llm-api-config.yml \
188
+ --port 8000 \
189
+ <YOUR_MODEL_DIR>
190
+ ```
191
+
192
+ ## Others
193
+
194
+ Kimi-K2 reuses the `DeepSeekV3CausalLM` architecture and convert it's weight into proper shape to save redevelopment effort. To let inference engines distinguish it from DeepSeek-V3 and apply the best optimizations, we set `"model_type": "kimi_k2"` in `config.json`.
195
+
196
+ If you are using a framework that is not on the recommended list, you can still run the model by manually changing `model_type` to "deepseek_v3" in `config.json` as a temporary workaround. You may need to manually parse tool calls in case no tool call parser is available in your framework.
docs/tool_call_guidance.md ADDED
@@ -0,0 +1,241 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ## Tool Calling
2
+ To enable the tool calling feature, you may need to set certain tool calling parser options when starting the service. See [deploy_guidance](./deploy_guidance.md) for details.
3
+ In Kimi-K2, a tool calling process includes:
4
+ - Passing function descriptions to Kimi-K2
5
+ - Kimi-K2 decides to make a function call and returns the necessary information for the function call to the user
6
+ - The user performs the function call, collects the call results, and passes the function call results to Kimi-K2
7
+ - Kimi-K2 continues to generate content based on the function call results until the model believes it has obtained sufficient information to respond to the user
8
+
9
+ ### Preparing Tools
10
+ Suppose we have a function `get_weather` that can query the weather conditions in real-time.
11
+ This function accepts a city name as a parameter and returns the weather conditions. We need to prepare a structured description for it so that Kimi-K2 can understand its functionality.
12
+
13
+ ```python
14
+ def get_weather(city):
15
+ return {"weather": "Sunny"}
16
+
17
+ # Collect the tool descriptions in tools
18
+ tools = [{
19
+ "type": "function",
20
+ "function": {
21
+ "name": "get_weather",
22
+ "description": "Get weather information. Call this tool when the user needs to get weather information",
23
+ "parameters": {
24
+ "type": "object",
25
+ "required": ["city"],
26
+ "properties": {
27
+ "city": {
28
+ "type": "string",
29
+ "description": "City name",
30
+ }
31
+ }
32
+ }
33
+ }
34
+ }]
35
+
36
+ # Tool name->object mapping for easy calling later
37
+ tool_map = {
38
+ "get_weather": get_weather
39
+ }
40
+ ```
41
+ ### Chat with tools
42
+ We use `openai.OpenAI` to send messages to Kimi-K2 along with tool descriptions. Kimi-K2 will autonomously decide whether to use and how to use the provided tools.
43
+ If Kimi-K2 believes a tool call is needed, it will return a result with `finish_reason='tool_calls'`. At this point, the returned result includes the tool call information.
44
+ After calling tools with the provided information, we then need to append the tool call results to the chat history and continue calling Kimi-K2.
45
+ Kimi-K2 may need to call tools multiple times until the model believes the current results can answer the user's question. We should check `finish_reason` until it is not `tool_calls`.
46
+
47
+ The results obtained by the user after calling the tools should be added to `messages` with `role='tool'`.
48
+
49
+ ```python
50
+ import json
51
+ from openai import OpenAI
52
+ model_name='moonshotai/Kimi-K2-Instruct'
53
+ client = OpenAI(base_url=endpoint,
54
+ api_key='xxx')
55
+
56
+ messages = [
57
+ {"role": "user", "content": "What's the weather like in Beijing today? Let's check using the tool."}
58
+ ]
59
+ finish_reason = None
60
+ while finish_reason is None or finish_reason == "tool_calls":
61
+ completion = client.chat.completions.create(
62
+ model=model_name,
63
+ messages=messages,
64
+ temperature=0.3,
65
+ tools=tools,
66
+ tool_choice="auto",
67
+ )
68
+ choice = completion.choices[0]
69
+ finish_reason = choice.finish_reason
70
+ # Note: The finish_reason when tool calls end may vary across different engines, so this condition check needs to be adjusted accordingly
71
+ if finish_reason == "tool_calls":
72
+ messages.append(choice.message)
73
+ for tool_call in choice.message.tool_calls:
74
+ tool_call_name = tool_call.function.name
75
+ tool_call_arguments = json.loads(tool_call.function.arguments)
76
+ tool_function = tool_map[tool_call_name]
77
+ tool_result = tool_function(tool_call_arguments)
78
+ print("tool_result", tool_result)
79
+
80
+ messages.append({
81
+ "role": "tool",
82
+ "tool_call_id": tool_call.id,
83
+ "name": tool_call_name,
84
+ "content": json.dumps(tool_result),
85
+ })
86
+ print('-' * 100)
87
+ print(choice.message.content)
88
+ ```
89
+ ### Tool Calling in Streaming Mode
90
+ Tool calling can also be used in streaming mode. In this case, we need to collect the tool call information returned in the stream until we have a complete tool call. Please refer to the code below:
91
+
92
+ ```python
93
+ messages = [
94
+ {"role": "user", "content": "What's the weather like in Beijing today? Let's check using the tool."}
95
+ ]
96
+ finish_reason = None
97
+ msg = ''
98
+ while finish_reason is None or finish_reason == "tool_calls":
99
+ completion = client.chat.completions.create(
100
+ model=model_name,
101
+ messages=messages,
102
+ temperature=0.3,
103
+ tools=tools,
104
+ tool_choice="auto",
105
+ stream=True
106
+ )
107
+ tool_calls = []
108
+ for chunk in completion:
109
+ delta = chunk.choices[0].delta
110
+ if delta.content:
111
+ msg += delta.content
112
+ if delta.tool_calls:
113
+ for tool_call_chunk in delta.tool_calls:
114
+ if tool_call_chunk.index is not None:
115
+ # Extend the tool_calls list
116
+ while len(tool_calls) <= tool_call_chunk.index:
117
+ tool_calls.append({
118
+ "id": "",
119
+ "type": "function",
120
+ "function": {
121
+ "name": "",
122
+ "arguments": ""
123
+ }
124
+ })
125
+
126
+ tc = tool_calls[tool_call_chunk.index]
127
+
128
+ if tool_call_chunk.id:
129
+ tc["id"] += tool_call_chunk.id
130
+ if tool_call_chunk.function.name:
131
+ tc["function"]["name"] += tool_call_chunk.function.name
132
+ if tool_call_chunk.function.arguments:
133
+ tc["function"]["arguments"] += tool_call_chunk.function.arguments
134
+
135
+ finish_reason = chunk.choices[0].finish_reason
136
+ # Note: The finish_reason when tool calls end may vary across different engines, so this condition check needs to be adjusted accordingly
137
+ if finish_reason == "tool_calls":
138
+ for tool_call in tool_calls:
139
+ tool_call_name = tool_call['function']['name']
140
+ tool_call_arguments = json.loads(tool_call['function']['arguments'])
141
+ tool_function = tool_map[tool_call_name]
142
+ tool_result = tool_function(tool_call_arguments)
143
+ messages.append({
144
+ "role": "tool",
145
+ "tool_call_id": tool_call['id'],
146
+ "name": tool_call_name,
147
+ "content": json.dumps(tool_result),
148
+ })
149
+ # The text generated by the tool call is not the final version, reset msg
150
+ msg = ''
151
+
152
+ print(msg)
153
+ ```
154
+ ### Manually Parsing Tool Calls
155
+ The tool call requests generated by Kimi-K2 can also be parsed manually, which is especially useful when the service you are using does not provide a tool-call parser.
156
+ The tool call requests generated by Kimi-K2 are wrapped by `<|tool_calls_section_begin|>` and `<|tool_calls_section_end|>`,
157
+ with each tool call wrapped by `<|tool_call_begin|>` and `<|tool_call_end|>`. The tool ID and arguments are separated by `<|tool_call_argument_begin|>`.
158
+ The format of the tool ID is `functions.{func_name}:{idx}`, from which we can parse the function name.
159
+
160
+ Based on the above rules, we can directly post request to the completions interface and manually parse tool calls.
161
+
162
+ ```python
163
+ import requests
164
+ from transformers import AutoTokenizer
165
+ messages = [
166
+ {"role": "user", "content": "What's the weather like in Beijing today? Let's check using the tool."}
167
+ ]
168
+ msg = ''
169
+ tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
170
+ while True:
171
+ text = tokenizer.apply_chat_template(
172
+ messages,
173
+ tokenize=False,
174
+ tools=tools,
175
+ add_generation_prompt=True,
176
+ )
177
+ payload = {
178
+ "model": model_name,
179
+ "prompt": text,
180
+ "max_tokens": 512
181
+ }
182
+ response = requests.post(
183
+ f"{endpoint}/completions",
184
+ headers={"Content-Type": "application/json"},
185
+ json=payload,
186
+ stream=False,
187
+ )
188
+ raw_out = response.json()
189
+
190
+ raw_output = raw_out["choices"][0]["text"]
191
+ tool_calls = extract_tool_call_info(raw_output)
192
+ if len(tool_calls) == 0:
193
+ # No tool calls
194
+ msg = raw_output
195
+ break
196
+ else:
197
+ for tool_call in tool_calls:
198
+ tool_call_name = tool_call['function']['name']
199
+ tool_call_arguments = json.loads(tool_call['function']['arguments'])
200
+ tool_function = tool_map[tool_call_name]
201
+ tool_result = tool_function(tool_call_arguments)
202
+
203
+ messages.append({
204
+ "role": "tool",
205
+ "tool_call_id": tool_call['id'],
206
+ "name": tool_call_name,
207
+ "content": json.dumps(tool_result),
208
+ })
209
+ print('-' * 100)
210
+ print(msg)
211
+ ```
212
+ Here, `extract_tool_call_info` parses the model output and returns the model call information. A simple implementation would be:
213
+ ```python
214
+ def extract_tool_call_info(tool_call_rsp: str):
215
+ if '<|tool_calls_section_begin|>' not in tool_call_rsp:
216
+ # No tool calls
217
+ return []
218
+ import re
219
+ pattern = r"<\|tool_calls_section_begin\|>(.*?)<\|tool_calls_section_end\|>"
220
+
221
+ tool_calls_sections = re.findall(pattern, tool_call_rsp, re.DOTALL)
222
+
223
+ # Extract multiple tool calls
224
+ func_call_pattern = r"<\|tool_call_begin\|>\s*(?P<tool_call_id>[\w\.]+:\d+)\s*<\|tool_call_argument_begin\|>\s*(?P<function_arguments>.*?)\s*<\|tool_call_end\|>"
225
+ tool_calls = []
226
+ for match in re.findall(func_call_pattern, tool_calls_sections[0], re.DOTALL):
227
+ function_id, function_args = match
228
+ # function_id: functions.get_weather:0
229
+ function_name = function_id.split('.')[1].split(':')[0]
230
+ tool_calls.append(
231
+ {
232
+ "id": function_id,
233
+ "type": "function",
234
+ "function": {
235
+ "name": function_name,
236
+ "arguments": function_args
237
+ }
238
+ }
239
+ )
240
+ return tool_calls
241
+ ```
figures/Base-Evaluation.png ADDED

Git LFS Details

  • SHA256: d1d3ee49430417c17326c9def19264756a3bc0b0aa001e598d0e0d751ebf93f8
  • Pointer size: 131 Bytes
  • Size of remote file: 245 kB
figures/kimi-logo.png ADDED