kimi source tree.txt · Lines 6001–6675

6001<td align="center" style="vertical-align: middle">88.5</td> 6002<td align="center" style="vertical-align: middle">-</td> 6003<td align="center" style="vertical-align: middle">82.0*</td> 6004</tr> 6005<tr> 6006<td align="center" style="vertical-align: middle">InfoVQA (val)</td> 6007<td align="center" style="vertical-align: middle">92.6</td> 6008<td align="center" style="vertical-align: middle">84*</td> 6009<td align="center" style="vertical-align: middle">76.9*</td> 6010<td align="center" style="vertical-align: middle">57.2*</td> 6011<td align="center" style="vertical-align: middle">-</td> 6012<td align="center" style="vertical-align: middle">89.5</td> 6013</tr> 6014<tr> 6015<td align="center" style="vertical-align: middle">SimpleVQA</td> 6016<td align="center" style="vertical-align: middle">71.2</td> 6017<td align="center" style="vertical-align: middle">55.8*</td> 6018<td align="center" style="vertical-align: middle">69.7*</td> 6019<td align="center" style="vertical-align: middle">69.7*</td> 6020<td align="center" style="vertical-align: middle">-</td> 6021<td align="center" style="vertical-align: middle">56.8*</td> 6022</tr> 6023<tr> 6024<td align="center" style="vertical-align: middle"><a href="https://github.com/MoonshotAI/WorldVQA">WorldVQA</a></td> 6025<td align="center" style="vertical-align: middle">46.3</td> 6026<td align="center" style="vertical-align: middle">28.0</td> 6027<td align="center" style="vertical-align: middle">36.8</td> 6028<td align="center" style="vertical-align: middle">47.4</td> 6029<td align="center" style="vertical-align: middle">-</td> 6030<td align="center" style="vertical-align: middle">23.5</td> 6031</tr> 6032<tr> 6033<td align="center" style="vertical-align: middle">VideoMMMU</td> 6034<td align="center" style="vertical-align: middle">86.6</td> 6035<td align="center" style="vertical-align: middle">85.9</td> 6036<td align="center" style="vertical-align: middle">84.4*</td> 6037<td align="center" style="vertical-align: middle">87.6</td> 6038<td align="center" style="vertical-align: middle">-</td> 6039<td align="center" style="vertical-align: middle">80.0</td> 6040</tr> 6041<tr> 6042<td align="center" style="vertical-align: middle">MMVU</td> 6043<td align="center" style="vertical-align: middle">80.4</td> 6044<td align="center" style="vertical-align: middle">80.8*</td> 6045<td align="center" style="vertical-align: middle">77.3</td> 6046<td align="center" style="vertical-align: middle">77.5</td> 6047<td align="center" style="vertical-align: middle">-</td> 6048<td align="center" style="vertical-align: middle">71.1</td> 6049</tr> 6050<tr> 6051<td align="center" style="vertical-align: middle">MotionBench</td> 6052<td align="center" style="vertical-align: middle">70.4</td> 6053<td align="center" style="vertical-align: middle">64.8</td> 6054<td align="center" style="vertical-align: middle">60.3</td> 6055<td align="center" style="vertical-align: middle">70.3</td> 6056<td align="center" style="vertical-align: middle">-</td> 6057<td align="center" style="vertical-align: middle">-</td> 6058</tr> 6059<tr> 6060<td align="center" style="vertical-align: middle">VideoMME</td> 6061<td align="center" style="vertical-align: middle">87.4</td> 6062<td align="center" style="vertical-align: middle">86.0*</td> 6063<td align="center" style="vertical-align: middle">-</td> 6064<td align="center" style="vertical-align: middle">88.4*</td> 6065<td align="center" style="vertical-align: middle">-</td> 6066<td align="center" style="vertical-align: middle">79.0</td> 6067</tr> 6068<tr> 6069<td align="center" style="vertical-align: middle">LongVideoBench</td> 6070<td align="center" style="vertical-align: middle">79.8</td> 6071<td align="center" style="vertical-align: middle">76.5*</td> 6072<td align="center" style="vertical-align: middle">67.2*</td> 6073<td align="center" style="vertical-align: middle">77.7*</td> 6074<td align="center" style="vertical-align: middle">-</td> 6075<td align="center" style="vertical-align: middle">65.6*</td> 6076</tr> 6077<tr> 6078<td align="center" style="vertical-align: middle">LVBench</td> 6079<td align="center" style="vertical-align: middle">75.9</td> 6080<td align="center" style="vertical-align: middle">-</td> 6081<td align="center" style="vertical-align: middle">-</td> 6082<td align="center" style="vertical-align: middle">73.5*</td> 6083<td align="center" style="vertical-align: middle">-</td> 6084<td align="center" style="vertical-align: middle">63.6</td> 6085</tr> 6086<tr> 6087<td align="center" colspan=8><strong>Coding</strong></td> 6088</tr> 6089<tr> 6090<td align="center" style="vertical-align: middle">SWE-Bench Verified</td> 6091<td align="center" style="vertical-align: middle">76.8</td> 6092<td align="center" style="vertical-align: middle">80.0</td> 6093<td align="center" style="vertical-align: middle">80.9</td> 6094<td align="center" style="vertical-align: middle">76.2</td> 6095<td align="center" style="vertical-align: middle">73.1</td> 6096<td align="center" style="vertical-align: middle">-</td> 6097</tr> 6098<tr> 6099<td align="center" style="vertical-align: middle">SWE-Bench Pro</td> 6100<td align="center" style="vertical-align: middle">50.7</td> 6101<td align="center" style="vertical-align: middle">55.6</td> 6102<td align="center" style="vertical-align: middle">55.4*</td> 6103<td align="center" style="vertical-align: middle">-</td> 6104<td align="center" style="vertical-align: middle">-</td> 6105<td align="center" style="vertical-align: middle">-</td> 6106</tr> 6107<tr> 6108<td align="center" style="vertical-align: middle">SWE-Bench Multilingual</td> 6109<td align="center" style="vertical-align: middle">73.0</td> 6110<td align="center" style="vertical-align: middle">72.0</td> 6111<td align="center" style="vertical-align: middle">77.5</td> 6112<td align="center" style="vertical-align: middle">65.0</td> 6113<td align="center" style="vertical-align: middle">70.2</td> 6114<td align="center" style="vertical-align: middle">-</td> 6115</tr> 6116<tr> 6117<td align="center" style="vertical-align: middle">Terminal Bench 2.0</td> 6118<td align="center" style="vertical-align: middle">50.8</td> 6119<td align="center" style="vertical-align: middle">54.0</td> 6120<td align="center" style="vertical-align: middle">59.3</td> 6121<td align="center" style="vertical-align: middle">54.2</td> 6122<td align="center" style="vertical-align: middle">46.4</td> 6123<td align="center" style="vertical-align: middle">-</td> 6124</tr> 6125<tr> 6126<td align="center" style="vertical-align: middle">PaperBench</td> 6127<td align="center" style="vertical-align: middle">63.5</td> 6128<td align="center" style="vertical-align: middle">63.7*</td> 6129<td align="center" style="vertical-align: middle">72.9*</td> 6130<td align="center" style="vertical-align: middle">-</td> 6131<td align="center" style="vertical-align: middle">47.1</td> 6132<td align="center" style="vertical-align: middle">-</td> 6133</tr> 6134<tr> 6135<td align="center" style="vertical-align: middle">CyberGym</td> 6136<td align="center" style="vertical-align: middle">41.3</td> 6137<td align="center" style="vertical-align: middle">-</td> 6138<td align="center" style="vertical-align: middle">50.6</td> 6139<td align="center" style="vertical-align: middle">39.9*</td> 6140<td align="center" style="vertical-align: middle">17.3*</td> 6141<td align="center" style="vertical-align: middle">-</td> 6142</tr> 6143<tr> 6144<td align="center" style="vertical-align: middle">SciCode</td> 6145<td align="center" style="vertical-align: middle">48.7</td> 6146<td align="center" style="vertical-align: middle">52.1</td> 6147<td align="center" style="vertical-align: middle">49.5</td> 6148<td align="center" style="vertical-align: middle">56.1</td> 6149<td align="center" style="vertical-align: middle">38.9</td> 6150<td align="center" style="vertical-align: middle">-</td> 6151</tr> 6152<tr> 6153<td align="center" style="vertical-align: middle">OJBench (cpp)</td> 6154<td align="center" style="vertical-align: middle">57.4</td> 6155<td align="center" style="vertical-align: middle">-</td> 6156<td align="center" style="vertical-align: middle">54.6*</td> 6157<td align="center" style="vertical-align: middle">68.5*</td> 6158<td align="center" style="vertical-align: middle">54.7*</td> 6159<td align="center" style="vertical-align: middle">-</td> 6160</tr> 6161<tr> 6162<td align="center" style="vertical-align: middle">LiveCodeBench (v6)</td> 6163<td align="center" style="vertical-align: middle">85.0</td> 6164<td align="center" style="vertical-align: middle">-</td> 6165<td align="center" style="vertical-align: middle">82.2*</td> 6166<td align="center" style="vertical-align: middle">87.4*</td> 6167<td align="center" style="vertical-align: middle">83.3</td> 6168<td align="center" style="vertical-align: middle">-</td> 6169</tr> 6170<tr> 6171<td align="center" colspan=8><strong>Long Context</strong></td> 6172</tr> 6173<tr> 6174<td align="center" style="vertical-align: middle">Longbench v2</td> 6175<td align="center" style="vertical-align: middle">61.0</td> 6176<td align="center" style="vertical-align: middle">54.5*</td> 6177<td align="center" style="vertical-align: middle">64.4*</td> 6178<td align="center" style="vertical-align: middle">68.2*</td> 6179<td align="center" style="vertical-align: middle">59.8*</td> 6180<td align="center" style="vertical-align: middle">-</td> 6181</tr> 6182<tr> 6183<td align="center" style="vertical-align: middle">AA-LCR</td> 6184<td align="center" style="vertical-align: middle">70.0</td> 6185<td align="center" style="vertical-align: middle">72.3*</td> 6186<td align="center" style="vertical-align: middle">71.3*</td> 6187<td align="center" style="vertical-align: middle">65.3*</td> 6188<td align="center" style="vertical-align: middle">64.3*</td> 6189<td align="center" style="vertical-align: middle">-</td> 6190<tr> 6191<td align="center" colspan=8><strong>Agentic Search</strong></td> 6192</tr> 6193<tr> 6194<td align="center" style="vertical-align: middle">BrowseComp</td> 6195<td align="center" style="vertical-align: middle">60.6</td> 6196<td align="center" style="vertical-align: middle" rowspan="2">65.8</td> 6197<td align="center" style="vertical-align: middle">37.0</td> 6198<td align="center" style="vertical-align: middle">37.8</td> 6199<td align="center" style="vertical-align: middle">51.4</td> 6200<td align="center" style="vertical-align: middle">-</td> 6201</tr> 6202<tr> 6203<td align="center" style="vertical-align: middle">BrowseComp<br>(w/ctx manage)</td> 6204<td align="center" style="vertical-align: middle">74.9</td> 6205<td align="center" style="vertical-align: middle">57.8</td> 6206<td align="center" style="vertical-align: middle">59.2</td> 6207<td align="center" style="vertical-align: middle">67.6</td> 6208<td align="center" style="vertical-align: middle">-</td> 6209</tr> 6210<tr> 6211<td align="center" style="vertical-align: middle">BrowseComp<br>(Agent Swarm)</td> 6212<td align="center" style="vertical-align: middle">78.4</td> 6213<td align="center" style="vertical-align: middle">-</td> 6214<td align="center" style="vertical-align: middle">-</td> 6215<td align="center" style="vertical-align: middle">-</td> 6216<td align="center" style="vertical-align: middle">-</td> 6217<td align="center" style="vertical-align: middle">-</td> 6218</tr> 6219<tr> 6220<td align="center" style="vertical-align: middle">WideSearch<br> (item-f1)</td> 6221<td align="center" style="vertical-align: middle">72.7</td> 6222<td align="center" style="vertical-align: middle">-</td> 6223<td align="center" style="vertical-align: middle">76.2*</td> 6224<td align="center" style="vertical-align: middle">57.0</td> 6225<td align="center" style="vertical-align: middle">32.5*</td> 6226<td align="center" style="vertical-align: middle">-</td> 6227</tr> 6228<tr> 6229<td align="center" style="vertical-align: middle">WideSearch<br> (item-f1 Agent Swarm)</td> 6230<td align="center" style="vertical-align: middle">79.0</td> 6231<td align="center" style="vertical-align: middle">-</td> 6232<td align="center" style="vertical-align: middle">-</td> 6233<td align="center" style="vertical-align: middle">-</td> 6234<td align="center" style="vertical-align: middle">-</td> 6235<td align="center" style="vertical-align: middle">-</td> 6236</tr> 6237<tr> 6238<td align="center" style="vertical-align: middle">DeepSearchQA</td> 6239<td align="center" style="vertical-align: middle">77.1</td> 6240<td align="center" style="vertical-align: middle">71.3*</td> 6241<td align="center" style="vertical-align: middle">76.1*</td> 6242<td align="center" style="vertical-align: middle">63.2*</td> 6243<td align="center" style="vertical-align: middle">60.9*</td> 6244<td align="center" style="vertical-align: middle">-</td> 6245</tr> 6246<tr> 6247<td align="center" style="vertical-align: middle">FinSearchCompT2&T3</td> 6248<td align="center" style="vertical-align: middle">67.8</td> 6249<td align="center" style="vertical-align: middle">-</td> 6250<td align="center" style="vertical-align: middle">66.2*</td> 6251<td align="center" style="vertical-align: middle">49.9</td> 6252<td align="center" style="vertical-align: middle">59.1*</td> 6253<td align="center" style="vertical-align: middle">-</td> 6254</tr> 6255<tr> 6256<td align="center" style="vertical-align: middle">Seal-0</td> 6257<td align="center" style="vertical-align: middle">57.4</td> 6258<td align="center" style="vertical-align: middle">45.0</td> 6259<td align="center" style="vertical-align: middle">47.7*</td> 6260<td align="center" style="vertical-align: middle">45.5*</td> 6261<td align="center" style="vertical-align: middle">49.5*</td> 6262<td align="center" style="vertical-align: middle">-</td> 6263</tr> 6264</tbody> 6265</table> 6266</div> 6267 6268<details> 6269<summary><b>Footnotes</b></summary> 6270 62711. General Testing Details 6272 - We report results for Kimi K2.5 and DeepSeek-V3.2 with thinking mode enabled, Claude Opus 4.5 with extended thinking mode, GPT-5.2 with xhigh reasoning effort, and Gemini 3 Pro with a high thinking level. For vision benchmarks, we additionally report results for Qwen3-VL-235B-A22B-Thinking. 6273 - Unless otherwise specified, all Kimi K2.5 experiments were conducted with temperature = 1.0, top-p = 0.95, and a context length of 256k tokens. 6274 - Benchmarks without publicly available scores were re-evaluated under the same conditions used for Kimi K2.5 and are marked with an asterisk (*). 6275 - We could not evaluate GPT-5.2 xhigh on all benchmarks due to service stability issues. For benchmarks that were not tested, we mark them as "-". 62762. Text and Reasoning 6277 - HLE, AIME 2025, HMMT 2025 (Feb), and GPQA-Diamond were evaluated with a maximum completion budget of 96k tokens. 6278 - Results for AIME and HMMT are averaged over 32 runs (avg@32); GPQA-Diamond over 8 runs (avg@8). 6279 - For HLE, we report scores on the full set (text & image). Kimi K2.5 scores 31.5 (text) and 21.3 (image) without tools, and 51.8 (text) and 39.8 (image) with tools. The DeepSeek-V3.2 score corresponds to its text-only subset (marked with †) . Hugging Face access was blocked to prevent potential data leakage. HLE with tools uses simple context management: once the context exceeds a threshold, only the latest round of tool messages is retained. 62803. Tool-Augmented / Agentic Search 6281 - Kimi K2.5 was equipped with search, code-interpreter, and web-browsing tools for HLE with tools and all agentic search benchmarks. 6282 - Except for BrowseComp (where K2.5 and DeepSeek-V3.2 used the discard-all strategy), no context management was applied, and tasks exceeding the supported context length were directly counted as failed. 6283 - The test system prompts emphasize deep and proactive tool use, instructing models to reason carefully, leverage tools, and verify uncertain information. Full prompts will be provided in the technical report. 6284 - Results for Seal-0 and WideSearch are averaged over four runs (avg@4). 62854. Vision Benchmarks 6286 - Max-tokens = 64k, averaged over three runs (avg@3). 6287 - ZeroBench (w/ tools) uses max-tokens-per-step = 24k and max-steps = 30 for multi-step reasoning. 6288 - MMMU-Pro follows the official protocol, preserving input order and prepending images. 6289 - GPT-5.2-xhigh had ~10% failure rate (no output despite 3 retries), treated as incorrect; reported scores likely underestimate true performance. 6290 - WorldVQA, a benchmark designed to evaluate atomic vision-centric world knowledge. Access WorldVQA at https://github.com/MoonshotAI/WorldVQA. 6291 - OmniDocBench Score is computed as (1 − normalized Levenshtein distance) × 100, where a higher score denotes superior accuracy. 62925. Coding Tasks 6293 - Terminal-Bench 2.0 scores were obtained with the default agent framework (Terminus-2) and the provided JSON parser. In our implementation, we evaluated Terminal-Bench 2.0 under non-thinking mode. This choice was made because our current context management strategy for the thinking mode is incompatible with Terminus-2. 6294 - For the SWE-Bench series of evaluations (including verified, multilingual, and pro), we used an internally developed evaluation framework. This framework includes a minimal set of tools—bash tool, createfile tool, insert tool, view tool, strreplace tool, and submit tool—along with tailored system prompts designed for the tasks. The highest scores were achieved under non-thinking mode. 6295 - The score of Claude Opus 4.5 on CyberGym is reported under the non-thinking setting. 6296 - All reported scores of coding tasks are averaged over 5 independent runs. 62976. Long-Context Benchmarks 6298 - AA-LCR: scores averaged over three runs (avg@3). 6299 - LongBench-V2: identical prompts and input contexts standardized to ~128k tokens. 63007. Agent Swarm 6301 - BrowseComp (Swarm Mode): main agent max 15 steps; sub-agents max 100 steps. 6302 - WideSearch (Swarm Mode): main and sub-agents max 100 steps. 6303 6304</details> 6305 6306## 4. Native INT4 Quantization 6307Kimi-K2.5 adopts the same native int4 quantization method as [Kimi-K2-Thinking](https://huggingface.co/moonshotai/Kimi-K2-Thinking#4-native-int4-quantization). 6308 6309## 5. Deployment 6310> [!Note] 6311> You can access Kimi-K2.5's API on https://platform.moonshot.ai and we provide OpenAI/Anthropic-compatible API for you. To verify the deployment is correct, we also provide the [Kimi Vendor Verifier](https://kimi.com/blog/kimi-vendor-verifier.html). 6312Currently, Kimi-K2.5 is recommended to run on the following inference engines: 6313* vLLM 6314* SGLang 6315* KTransformers 6316 6317The minimum version requirement for `transformers` is `4.57.1`. 6318 6319Deployment examples can be found in the [Model Deployment Guide](docs/deploy_guidance.md). 6320 6321 6322--- 6323## 6. Model Usage 6324 6325The usage demos below demonstrate how to call our official API. 6326 6327For third-party APIs deployed with vLLM or SGLang, please note that: 6328> [!Note] 6329> - Chat with video content is an experimental feature and is only supported in our official API for now. 6330> 6331> - The recommended `temperature` will be `1.0` for Thinking mode and `0.6` for Instant mode. 6332> 6333> - The recommended `top_p` is `0.95`. 6334> 6335> - To use instant mode, you need to pass `{'chat_template_kwargs': {"thinking": False}}` in `extra_body`. 6336 6337### Chat Completion 6338 6339This is a simple chat completion script which shows how to call K2.5 API in Thinking and Instant modes. 6340 6341```python 6342import openai 6343import base64 6344import requests 6345def simple_chat(client: openai.OpenAI, model_name: str): 6346 messages = [ 6347 {'role': 'system', 'content': 'You are Kimi, an AI assistant created by Moonshot AI.'}, 6348 { 6349 'role': 'user', 6350 'content': [ 6351 {'type': 'text', 'text': 'which one is bigger, 9.11 or 9.9? think carefully.'} 6352 ], 6353 }, 6354 ] 6355 response = client.chat.completions.create( 6356 model=model_name, messages=messages, stream=False, max_tokens=4096 6357 ) 6358 print('====== Below is reasoning_content in Thinking Mode ======') 6359 print(f'reasoning content: {response.choices[0].message.reasoning_content}') 6360 print('====== Below is response in Thinking Mode ======') 6361 print(f'response: {response.choices[0].message.content}') 6362 6363 # To use instant mode, pass {"thinking" = {"type":"disabled"}} 6364 response = client.chat.completions.create( 6365 model=model_name, 6366 messages=messages, 6367 stream=False, 6368 max_tokens=4096, 6369 extra_body={'thinking': {'type': 'disabled'}}, # this is for official API 6370 # extra_body= {'chat_template_kwargs': {"thinking": False}} # this is for vLLM/SGLang 6371 ) 6372 print('====== Below is response in Instant Mode ======') 6373 print(f'response: {response.choices[0].message.content}') 6374``` 6375 6376 6377### Chat Completion with visual content 6378 6379K2.5 supports Image and Video input. 6380 6381The following example demonstrates how to call K2.5 API with image input: 6382 6383```python 6384import openai 6385import base64 6386import requests 6387 6388def chat_with_image(client: openai.OpenAI, model_name: str): 6389 url = 'https://huggingface.co/moonshotai/Kimi-K2.5/resolve/main/figures/kimi-logo.png' 6390 image_base64 = base64.b64encode(requests.get(url).content).decode() 6391 messages = [ 6392 { 6393 'role': 'user', 6394 'content': [ 6395 { 6396 'type': 'image_url', 6397 'image_url': {'url': f'data:image/png;base64, {image_base64}'}, 6398 }, 6399 {'type': 'text', 'text': 'Describe this image in detail.'}, 6400 ], 6401 } 6402 ] 6403 6404 response = client.chat.completions.create( 6405 model=model_name, messages=messages, stream=False, max_tokens=8192 6406 ) 6407 print('====== Below is reasoning_content in Thinking Mode ======') 6408 print(f'reasoning content: {response.choices[0].message.reasoning_content}') 6409 print('====== Below is response in Thinking Mode ======') 6410 print(f'response: {response.choices[0].message.content}') 6411 6412 # Also support instant mode if you pass {"thinking" = {"type":"disabled"}} 6413 response = client.chat.completions.create( 6414 model=model_name, 6415 messages=messages, 6416 stream=False, 6417 max_tokens=4096, 6418 extra_body={'thinking': {'type': 'disabled'}}, # this is for official API 6419 # extra_body= {'chat_template_kwargs': {"thinking": False}} # this is for vLLM/SGLang 6420 ) 6421 print('====== Below is response in Instant Mode ======') 6422 print(f'response: {response.choices[0].message.content}') 6423 6424 return response.choices[0].message.content 6425``` 6426 6427The following example demonstrates how to call K2.5 API with video input: 6428 6429```python 6430import openai 6431import base64 6432import requests 6433 6434def chat_with_video(client: openai.OpenAI, model_name:str): 6435 url = 'https://huggingface.co/moonshotai/Kimi-K2.5/resolve/main/figures/demo_video.mp4' 6436 video_base64 = base64.b64encode(requests.get(url).content).decode() 6437 messages = [ 6438 { 6439 "role": "user", 6440 "content": [ 6441 { 6442 "type": "video_url", 6443 "video_url": {"url": f"data:video/mp4;base64,{video_base64}"}, 6444 }, 6445 {"type": "text","text": "Describe the video in detail."}, 6446 ], 6447 } 6448 ] 6449 6450 response = client.chat.completions.create(model=model_name, messages=messages) 6451 print('====== Below is reasoning_content in Thinking Mode ======') 6452 print(f'reasoning content: {response.choices[0].message.reasoning_content}') 6453 print('====== Below is response in Thinking Mode ======') 6454 print(f'response: {response.choices[0].message.content}') 6455 6456 # Also support instant mode if pass {"thinking" = {"type":"disabled"}} 6457 response = client.chat.completions.create( 6458 model=model_name, 6459 messages=messages, 6460 stream=False, 6461 max_tokens=4096, 6462 extra_body={'thinking': {'type': 'disabled'}}, # this is for official API 6463 # extra_body= {'chat_template_kwargs': {"thinking": False}} # this is for vLLM/SGLang 6464 ) 6465 print('====== Below is response in Instant Mode ======') 6466 print(f'response: {response.choices[0].message.content}') 6467 return response.choices[0].message.content 6468``` 6469 6470### Interleaved Thinking and Multi-Step Tool Call 6471 6472K2.5 shares the same design of Interleaved Thinking and Multi-Step Tool Call as K2 Thinking. For usage example, please refer to the [K2 Thinking documentation](https://platform.moonshot.ai/docs/guide/use-kimi-k2-thinking-model#complete-example). 6473 6474 6475### Coding Agent Framework 6476 6477Kimi K2.5 works best with Kimi Code CLI as its agent framework — give it a try at https://www.kimi.com/code. 6478 6479 6480--- 6481 6482## 7. License 6483 6484Both the code repository and the model weights are released under the [Modified MIT License](LICENSE). 6485 6486--- 6487 6488## 8. Third Party Notices 6489 6490See [THIRD PARTY NOTICES](THIRD_PARTY_NOTICES.md) 6491 6492--- 6493 6494## 9. Contact Us 6495 6496If you have any questions, please reach out at [[contact email redacted]](mailto:[contact email redacted]). 6497 6498## 10. Reference 6499 6500If you find K2.5 useful for your research, please kindly cite K2.5 technical report as follows: 6501 6502```bibtex 6503@misc{kimiteam2026kimik25visualagentic, 6504 title={Kimi K2.5: Visual Agentic Intelligence}, 6505 author={Kimi Team and Tongtong Bai and Yifan Bai and Yiping Bao and S. H. Cai and Yuan Cao and Y. Charles and H. S. Che and Cheng Chen and Guanduo Chen and Huarong Chen and Jia Chen and Jiahao Chen and Jianlong Chen and Jun Chen and Kefan Chen and Liang Chen and Ruijue Chen and Xinhao Chen and Yanru Chen and Yanxu Chen and Yicun Chen and Yimin Chen and Yingjiang Chen and Yuankun Chen and Yujie Chen and Yutian Chen and Zhirong Chen and Ziwei Chen and Dazhi Cheng and Minghan Chu and Jialei Cui and Jiaqi Deng and Muxi Diao and Hao Ding and Mengfan Dong and Mengnan Dong and Yuxin Dong and Yuhao Dong and Angang Du and Chenzhuang Du and Dikang Du and Lingxiao Du and Yulun Du and Yu Fan and Shengjun Fang and Qiulin Feng and Yichen Feng and Garimugai Fu and Kelin Fu and Hongcheng Gao and Tong Gao and Yuyao Ge and Shangyi Geng and Chengyang Gong and Xiaochen Gong and Zhuoma Gongque and Qizheng Gu and Xinran Gu and Yicheng Gu and Longyu Guan and Yuanying Guo and Xiaoru Hao and Weiran He and Wenyang He and Yunjia He and Chao Hong and Hao Hu and Jiaxi Hu and Yangyang Hu and Zhenxing Hu and Ke Huang and Ruiyuan Huang and Weixiao Huang and Zhiqi Huang and Tao Jiang and Zhejun Jiang and Xinyi Jin and Yu Jing and Guokun Lai and Aidi Li and C. Li and Cheng Li and Fang Li and Guanghe Li and Guanyu Li and Haitao Li and Haoyang Li and Jia Li and Jingwei Li and Junxiong Li and Lincan Li and Mo Li and Weihong Li and Wentao Li and Xinhang Li and Xinhao Li and Yang Li and Yanhao Li and Yiwei Li and Yuxiao Li and Zhaowei Li and Zheming Li and Weilong Liao and Jiawei Lin and Xiaohan Lin and Zhishan Lin and Zichao Lin and Cheng Liu and Chenyu Liu and Hongzhang Liu and Liang Liu and Shaowei Liu and Shudong Liu and Shuran Liu and Tianwei Liu and Tianyu Liu and Weizhou Liu and Xiangyan Liu and Yangyang Liu and Yanming Liu and Yibo Liu and Yuanxin Liu and Yue Liu and Zhengying Liu and Zhongnuo Liu and Enzhe Lu and Haoyu Lu and Zhiyuan Lu and Junyu Luo and Tongxu Luo and Yashuo Luo and Long Ma and Yingwei Ma and Shaoguang Mao and Yuan Mei and Xin Men and Fanqing Meng and Zhiyong Meng and Yibo Miao and Minqing Ni and Kun Ouyang and Siyuan Pan and Bo Pang and Yuchao Qian and Ruoyu Qin and Zeyu Qin and Jiezhong Qiu and Bowen Qu and Zeyu Shang and Youbo Shao and Tianxiao Shen and Zhennan Shen and Juanfeng Shi and Lidong Shi and Shengyuan Shi and Feifan Song and Pengwei Song and Tianhui Song and Xiaoxi Song and Hongjin Su and Jianlin Su and Zhaochen Su and Lin Sui and Jinsong Sun and Junyao Sun and Tongyu Sun and Flood Sung and Yunpeng Tai and Chuning Tang and Heyi Tang and Xiaojuan Tang and Zhengyang Tang and Jiawen Tao and Shiyuan Teng and Chaoran Tian and Pengfei Tian and Ao Wang and Bowen Wang and Chensi Wang and Chuang Wang and Congcong Wang and Dingkun Wang and Dinglu Wang and Dongliang Wang and Feng Wang and Hailong Wang and Haiming Wang and Hengzhi Wang and Huaqing Wang and Hui Wang and Jiahao Wang and Jinhong Wang and Jiuzheng Wang and Kaixin Wang and Linian Wang and Qibin Wang and Shengjie Wang and Shuyi Wang and Si Wang and Wei Wang and Xiaochen Wang and Xinyuan Wang and Yao Wang and Yejie Wang and Yipu Wang and Yiqin Wang and Yucheng Wang and Yuzhi Wang and Zhaoji Wang and Zhaowei Wang and Zhengtao Wang and Zhexu Wang and Zihan Wang and Zizhe Wang and Chu Wei and Ming Wei and Chuan Wen and Zichen Wen and Chengjie Wu and Haoning Wu and Junyan Wu and Rucong Wu and Wenhao Wu and Yuefeng Wu and Yuhao Wu and Yuxin Wu and Zijian Wu and Chenjun Xiao and Jin Xie and Xiaotong Xie and Yuchong Xie and Yifei Xin and Bowei Xing and Boyu Xu and Jianfan Xu and Jing Xu and Jinjing Xu and L. H. Xu and Lin Xu and Suting Xu and Weixin Xu and Xinbo Xu and Xinran Xu and Yangchuan Xu and Yichang Xu and Yuemeng Xu and Zelai Xu and Ziyao Xu and Junjie Yan and Yuzi Yan and Guangyao Yang and Hao Yang and Junwei Yang and Kai Yang and Ningyuan Yang and Ruihan Yang and Xiaofei Yang and Xinlong Yang and Ying Yang and Yi Yang and Yi Yang and Zhen Yang and Zhilin Yang and Zonghan Yang and Haotian Yao and Dan Ye and Wenjie Ye and Zhuorui Ye and Bohong Yin and Chengzhen Yu and Longhui Yu and Tao Yu and Tianxiang Yu and Enming Yuan and Mengjie Yuan and Xiaokun Yuan and Yang Yue and Weihao Zeng and Dunyuan Zha and Haobing Zhan and Dehao Zhang and Hao Zhang and Jin Zhang and Puqi Zhang and Qiao Zhang and Rui Zhang and Xiaobin Zhang and Y. Zhang and Yadong Zhang and Yangkun Zhang and Yichi Zhang and Yizhi Zhang and Yongting Zhang and Yu Zhang and Yushun Zhang and Yutao Zhang and Yutong Zhang and Zheng Zhang and Chenguang Zhao and Feifan Zhao and Jinxiang Zhao and Shuai Zhao and Xiangyu Zhao and Yikai Zhao and Zijia Zhao and Huabin Zheng and Ruihan Zheng and Shaojie Zheng and Tengyang Zheng and Junfeng Zhong and Longguang Zhong and Weiming Zhong and M. Zhou and Runjie Zhou and Xinyu Zhou and Zaida Zhou and Jinguo Zhu and Liya Zhu and Xinhao Zhu and Yuxuan Zhu and Zhen Zhu and Jingze Zhuang and Weiyu Zhuang and Ying Zou and Xinxing Zu}, 6506 year={2026}, 6507 eprint={2602.02276}, 6508 archivePrefix={arXiv}, 6509 primaryClass={cs.CL}, 6510 url={https://arxiv.org/abs/2602.02276}, 6511} 6512``` 6513 6514 6515>>>>> FILE: docs/deploy_guidance.md <<<<< 6516# Kimi-K2.5 Deployment Guide 6517 6518> [!Note] 6519> This guide only provides some examples of deployment commands for Kimi-K2.5, 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. 6520 6521> kimi_k2 reasoning parser and other related features have been merged into vLLM/sglang and will be available in the next release. For now, please use the nightly build Docker image. 6522## vLLM Deployment 6523 6524This model is available in nightly vLLM wheel: 6525``` 6526uv pip install -U vllm \ 6527 --torch-backend=auto \ 6528 --extra-index-url https://wheels.vllm.ai/nightly 6529``` 6530 6531Here is the example to serve this model on a H200 single node with TP8 via vLLM: 6532```bash 6533vllm serve $MODEL_PATH -tp 8 --mm-encoder-tp-mode data --trust-remote-code --tool-call-parser kimi_k2 --reasoning-parser kimi_k2 6534``` 6535**Key notes** 6536- `--tool-call-parser kimi_k2`: Required for enabling tool calling 6537- `--reasoning-parser kimi_k2`: Kimi-K2.5 enables thinking mode by default. Make sure to pass this for correct reasoning processing. 6538 6539## SGLang Deployment 6540 6541This model is available in SGLang latest main: 6542 6543``` 6544pip install "sglang @ git+https://github.com/sgl-project/sglang.git#subdirectory=python" 6545pip install nvidia-cudnn-cu12==9.16.0.29 6546``` 6547 6548Similarly, here is the example for it to run with TP8 on H200 in a single node via SGLang: 6549``` bash 6550sglang serve --model-path $MODEL_PATH --tp 8 --trust-remote-code --tool-call-parser kimi_k2 --reasoning-parser kimi_k2 6551``` 6552**Key parameter notes:** 6553- `--tool-call-parser kimi_k2`: Required when enabling tool usage. 6554- `--reasoning-parser kimi_k2`: Required for correctly processing reasoning content. 6555 6556## KTransformers Deployment 6557### KTransformers+SGLang Inference Deployment 6558Launch with KTransformers + SGLang for CPU+GPU heterogeneous inference: 6559 6560``` 6561python -m sglang.launch_server \ 6562 --model path/to/Kimi-K2.5/ \ 6563 --kt-amx-weight-path path/to/Kimi-K2.5/ \ 6564 --kt-cpuinfer 64 \ 6565 --kt-threadpool-count 2 \ 6566 --kt-num-gpu-experts 180 \ 6567 --kt-amx-method AMXINT4 \ 6568 --trust-remote-code \ 6569 --mem-fraction-static 0.98 \ 6570 --chunked-prefill-size 16384 \ 6571 --max-running-requests 48 \ 6572 --max-total-tokens 50000 \ 6573 --tensor-parallel-size 8 \ 6574 --enable-p2p-check \ 6575 --disable-shared-experts-fusion 6576``` 6577 6578Achieves 640.12 tokens/s Prefill and 24.51 tokens/s Decode (48-way concurrency) on 8× NVIDIA L20 + 2× Intel 6454S. 6579 6580More details: https://github.com/kvcache-ai/ktransformers/blob/main/doc/en/Kimi-K2.5.md . 6581 6582### KTransformers+LLaMA-Factory Fine-tuning Deployment 6583 6584You can use below command to run LoRA SFT with KT+llamafactory. 6585 6586``` 6587# For LoRA SFT 6588USE_KT=1 llamafactory-cli train examples/train_lora/kimik2_lora_sft_kt.yaml 6589# For Chat with model after LoRA SFT 6590llamafactory-cli chat examples/inference/kimik2_lora_sft_kt.yaml 6591# For API with model after LoRA SFT 6592llamafactory-cli api examples/inference/kimik2_lora_sft_kt.yaml 6593``` 6594 6595This achieves end-to-end LoRA SFT Throughput: 44.55 token/s on 2× NVIDIA 4090 + Intel 8488C with 1.97T RAM and 200G swap memory. 6596 6597More details refer to https://github.com/kvcache-ai/ktransformers/blob/main/doc/en/SFT_Installation_Guide_KimiK2.5.md . 6598 6599 6600>>>>> FILE: LICENSE <<<<< 6601Modified MIT License 6602 6603Copyright (c) 2026 Moonshot AI 6604 6605Permission is hereby granted, free of charge, to any person obtaining a copy 6606of this software and associated documentation files (the “Software”), to deal 6607in the Software without restriction, including without limitation the rights 6608to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 6609copies of the Software, and to permit persons to whom the Software is 6610furnished to do so, subject to the following conditions: 6611 6612The above copyright notice and this permission notice shall be included in all 6613copies or substantial portions of the Software. 6614 6615THE SOFTWARE IS PROVIDED “AS IS”, WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 6616IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 6617FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 6618AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 6619LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 6620OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 6621SOFTWARE. 6622 6623Our only modification part is that, if the Software (or any derivative works 6624thereof) is used for any of your commercial products or services that have 6625more than 100 million monthly active users, or more than 20 million US dollars 6626(or equivalent in other currencies) in monthly revenue, you shall prominently 6627display "Kimi K2.5" on the user interface of such product or service. 6628 6629 6630>>>>> FILE: THIRD_PARTY_NOTICES.md <<<<< 6631# THIRD_PARTY_NOTICES 6632 6633This file lists third-party software contained in Kimi-K2.5 along with their licenses, in compliance with the redistribution clauses of those licenses. 6634 6635--- 6636 6637## 1. DeepSeek-V3 6638 6639Our model archietecture is DeepSeek-V3-like. Some of modeling codes are copied from the source repository. 6640 6641- **Source Repository** 6642 https://huggingface.co/deepseek-ai/DeepSeek-V3 6643 6644- **Files / Directories Used** 6645 - configuration_deepseek.py 6646 - modeling_deepseek.py 6647 6648- **License Type** 6649 MIT License 6650 6651- **Copyright Notice** 6652 Copyright (c) 2023 DeepSeek 6653 6654- **Full License Text** 6655``` 6656MIT License 6657Copyright (c) 2023 DeepSeek 6658Permission is hereby granted, free of charge, to any person obtaining a copy 6659of this software and associated documentation files (the "Software"), to deal 6660in the Software without restriction, including without limitation the rights 6661to use, copy, modify, merge, publish, distribute, sublicense, and/or sell 6662copies of the Software, and to permit persons to whom the Software is 6663furnished to do so, subject to the following conditions: 6664The above copyright notice and this permission notice shall be included in all 6665copies or substantial portions of the Software. 6666THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 6667IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 6668FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE 6669AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 6670LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 6671OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 6672SOFTWARE. 6673``` 6674 6675