--- license: other license_name: kimi-k3 library_name: transformers pipeline_tag: image-text-to-text tags: - vLLM - cubic-quantization - W2A8 - W3A8 - W4A8 - W2A16 - W3A16 - W4A16 - multimodal base_model: - moonshotai/Kimi-K3 base_model_relation: quantized --- # Kimi-K3-Cubic-2.5Bit Base model: [moonshotai/Kimi-K3](https://huggingface.co/moonshotai/Kimi-K3) **2.8T parameters · 2.5-bit experts · one 8 × H200 node · ~700K context**

Keep Open Quantized Models Accessible

Large model checkpoints are expensive to host, and Hugging Face storage is becoming the main bottleneck.

Your support helps keep quantized models free and easy to download for researchers, developers, students, and small teams without large GPU budgets.

Every contribution goes toward storage costs, helping more people experiment, evaluate, and build with open models.

Support us on Ko-fi

Kimi-K3-Cubic-2.5Bit is the first public showcase of **Cubic quantization**, exploring practical ultra-low-bit, high-concurrency LLM inference. Cubic uses parameterized, non-uniform levels that adapt to each weight group, designed to challenge conventional integer and floating-point formats at ultra-low precision while relaxing the usual need for small group sizes; most quantized layers in this model use group size 512. Technical report: [arXiv:2608.06763](https://arxiv.org/abs/2608.06763). - **57 output tok/s** — single-request short-context decode - **800–900+ output tok/s** — 64-way concurrency, without dSpark or other speculative decoding - **W1–W8 runtime kernels** — native SM80/86/89/90/90a/100/120 coverage The export is fully data-free: no calibration dataset or activation samples were used. The complete converter is included as [`quantize_k3.py`](./quantize_k3.py). The same packed checkpoint runs in either weight-only A16 or dynamic-A8 mode through the [QuantTrio vLLM Cubic runtime](https://github.com/QuantTrio/vllm-cubic). This checkpoint itself uses mixed W2, W3, and W4 expert weights with an audited effective width of **2.4986 bits**. ### 【Quantization Policy】 | Scope | Format | |-------|--------| | Dense layer 0 | Source dtype (BF16) | | MoE layers 1–3 routed experts | Cubic W3, group size 256 | | MoE layers 4–32 routed experts | Cubic W3, group size 512 | | MoE layers 33–91 routed experts | Cubic W2, group size 512 | | MoE layer 92 routed experts | Cubic W4, group size 512 | | Attention/KDA/MLA weights | Source dtype (BF16) | | Shared experts and ordinary MLP weights | Source dtype (BF16) | | Vision tower, multimodal projector, embeddings, norms, and LM head | Source dtype | `Loss` is the scale-free NRMSE of groupwise weight reconstruction, not a downstream task-accuracy score. | Format | Loss | Clipped values | |--------|-----:|---------------:| | W2, group size 512 | 0.451251 | 12.2307% | | W3, group size 256 | 0.179608 | 2.6971% | | W3, group size 512 | 0.181007 | 2.9948% | | W4, group size 512 | 0.043047 | 0.5635% | ### 【Dependencies / Installation】 - Runtime: [vLLM Cubic v0.26.1+cubic.20260805](https://github.com/QuantTrio/vllm-cubic/releases/tag/v0.26.1%2Bcubic.20260805) - Python: 3.10–3.14 - Platform: Linux x86-64, PyTorch 2.13.0, CUDA 13.0 - Verified: Python 3.12, 8 × H200, TP8 + EP Install: ```bash uv venv --python 3.12 .venv source .venv/bin/activate uv pip install \ 'https://github.com/QuantTrio/vllm-cubic/releases/download/v0.26.1%2Bcubic.20260805/vllm-0.26.1%2Bcubic.20260805-cp38-abi3-linux_x86_64.whl' ``` ### 【Quantization Script】 - Script: [`quantize_k3.py`](./quantize_k3.py) ([source](https://github.com/QuantTrio/vllm-cubic/blob/main/examples/quantization/quantize_k3.py)) - Input: original `moonshotai/Kimi-K3` MXFP4 checkpoint - Method: data-free; dynamic-A8 carrier correction on by default (`--disable-a8-correction` to disable) - Time: approximately **35–40 minutes on 8 × H200** with fast local storage - Dependencies: included with vLLM Cubic; standalone: `pip install torch safetensors regex` - Output: destination path must not already exist ```bash python -u quantize_k3.py \ --source __YOUR_PATH__/moonshotai/Kimi-K3 \ --output __YOUR_PATH__/Kimi-K3-Cubic-2.5Bit \ --devices cuda:0,cuda:1,cuda:2,cuda:3,cuda:4,cuda:5,cuda:6,cuda:7 ``` ### 【vLLM Startup Command】 The following command enables dynamic A8 Cubic execution. Change `VLLM_CUBIC_DYNAMIC_A8` to `0` to run the same checkpoint in weight-only A16 mode. ```bash export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 export VLLM_CUBIC_DYNAMIC_A8=1 vllm serve \ __YOUR_PATH__/Kimi-K3-Cubic-2.5Bit \ --served-model-name Kimi-K3-Cubic-2.5Bit \ --trust-remote-code \ --quantization cubic \ --kv-cache-dtype fp8_q16 \ --gpu-memory-utilization 0.985 \ --tensor-parallel-size 8 \ --enable-expert-parallel \ --mm-encoder-tp-mode data \ --max-model-len auto \ --max-num-seqs 128 \ --max-num-batched-tokens 2048 \ --enable-prefix-caching \ --enable-auto-tool-choice \ --tool-call-parser kimi_k3 \ --reasoning-parser kimi_k3 \ --host 0.0.0.0 \ --port 8000 \ --disable-uvicorn-access-log ``` Observed performance for this command on 8 × NVIDIA H200 141GB and 8 × NVIDIA H20 141GB (2026-08-05) is: ### 【docker with vLLM Startup Command】 ```bash docker pull swr.cn-north-4.myhuaweicloud.com/desaysv/vllm-cubic:0.26.1-cubic.20260805 export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 export VLLM_CUBIC_DYNAMIC_A8=1 export FLASHINFER_DISABLE_VERSION_CHECK=1 vllm serve \ __YOUR_PATH__/Kimi-K3-Cubic-2.5Bit \ --served-model-name Kimi-K3-Cubic-2.5Bit \ --trust-remote-code \ --quantization cubic \ --kv-cache-dtype fp8_q16 \ --gpu-memory-utilization 0.985 \ --tensor-parallel-size 8 \ --enable-expert-parallel \ --mm-encoder-tp-mode data \ --max-model-len auto \ --max-num-seqs 128 \ --max-num-batched-tokens 2048 \ --enable-prefix-caching \ --enable-auto-tool-choice \ --tool-call-parser kimi_k3 \ --reasoning-parser kimi_k3 \ --host 0.0.0.0 \ --port 8000 \ --disable-uvicorn-access-log ``` *skip deepgemm error ,its not effect start ### 【benchmark】10*8K/1K | Item | Observed value | |------|---------------:| | Total requests | 10 | | Successful requests | 10 | | Failed requests | 0 | | Unfinished requests | 0 | | Average concurrency | 10 | | Total duration | 209.75 s | | Total throughput | 464.11 tokens/s | | Input throughput | 684.34 tokens/s | | Output throughput | 51.07 tokens/s | | Average request latency | 204.23 s | | Average TTFT | 69,607.26 ms | | Average TPOT | 164.76 ms | | Average ITL | 95.25 ms | | TTFT (P50) | 63,297.56 ms | | TTFT (P90) | 128,423.96 ms | | TTFT (P99) | 128,423.96 ms | | ITL (P50) | 81.85 ms | | ITL (P90) | 178.56 ms | | ITL (P99) | 178.56 ms | | TPOT (P50) | 154.40 ms | | TPOT (P90) | 206.35 ms | | TPOT (P99) | 206.35 ms | | Latency (P50) | 204.10 s | | Latency (P90) | 208.45 s | | Latency (P99) | 208.45 s | | Input tokens per request | 8,088 | | Output tokens per request | 1,000 | ### 【benchmark】10*64K/3000 | Item | Observed value | |------|---------------:| | Total requests | 10 | | Successful requests | 10 | | Failed requests | 0 | | Unfinished requests | 0 | | Average concurrency | 9 | | Total duration | 1,461.29 s | | Total throughput | 519.15 tokens/s | | Input throughput | 513.63 tokens/s | | Output throughput | 23.21 tokens/s | | Average request latency | 1,387.96 s | | Average TTFT | 882,028.37 ms | | Average TPOT | 334.97 ms | | Average ITL | 107.57 ms | | TTFT (P50) | 604,938.80 ms | | TTFT (P90) | 1,352,678.48 ms | | TTFT (P99) | 1,352,678.48 ms | | ITL (P50) | 61.83 ms | | ITL (P90) | 313.74 ms | | ITL (P99) | 313.74 ms | | TPOT (P50) | 378.19 ms | | TPOT (P90) | 465.73 ms | | TPOT (P99) | 465.73 ms | | Latency (P50) | 1,394.02 s | | Latency (P90) | 1,460.10 s | | Latency (P99) | 1,460.10 s | | Input tokens per request | 64,088 | | Output tokens per request | 3,000 | | Item | Observed value | |------|---------------:| | Dynamic Cubic mode | A8 | | Parallel topology | TP8 + EP | | KV-cache dtype | `fp8_q16` | | Effective maximum model length | 728,640 tokens | | Single-request short-context decode | approximately 57 output tok/s | | 64-request aggregate decode | approximately 800–900+ output tok/s | | Speculative decoding | None; dSpark was not used | The throughput values are stable decode-window measurements from short-prompt, long-output text requests, rather than end-to-end request throughput. Prompt length, generated context length, sampling, concurrent prefill, GPU clocks, and other workloads can change the result. Because `--max-model-len auto` derives the limit from free memory measured during startup, the effective token count can also vary slightly between launches. `fp8_q16` is an optional query-preserving FP8 KV-cache mode from the Cubic runtime. It keeps attention queries in the model dtype and automatically selects a compatible attention backend. To use a conventional BF16 KV cache, replace it with `--kv-cache-dtype auto`; this reduces the available context length but provides the highest-precision KV storage. The first launch may spend several minutes compiling Triton kernels, calibrating Cubic kernel choices for the tensors and GPU type in use, and capturing CUDA graphs. Calibration results are persisted in the Triton cache and reused by later launches on compatible devices. ### 【Validation Status】 - 8 × H200 TP8 + EP model loading and text generation verified. - Weight-only A16 and dynamic-A8 execution verified with the same checkpoint. - `fp8_q16` and BF16 KV-cache serving paths verified. - No formal downstream benchmark is claimed for this quantized checkpoint; evaluate quality and performance for your own workloads before deployment. ### 【Model Files】 | Item | Value | |------|-------| | Model data | 899 GiB | | Safetensors shards | 341 | | Last updated | 2026-08-05 | ### 【Model Download】 ```python from huggingface_hub import snapshot_download snapshot_download( "QuantTrio/Kimi-K3-Cubic-2.5Bit", local_dir="your_local_path/Kimi-K3-Cubic-2.5Bit", ) ``` ### 【License】 This quantized checkpoint is derived from Kimi K3 and remains subject to the [Kimi K3 License](https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE). The QuantTrio vLLM Cubic runtime is distributed separately under its own repository license. ### 【Base Model Overview】 The following material is reproduced from the upstream Kimi K3 model card and describes the base model. Upstream MXFP4 deployment instructions do not apply to this Cubic checkpoint; use the Cubic installation and launch command above.
Kimi K3
Chat Homepage
Hugging Face Twitter Follow Discord ModelScope
License

📰  Tech Blog |     📄  Full Report

## 1. Model Introduction Kimi K3 is an open-weight, native multimodal agentic model and our most capable model to date. It is a 2.8T-parameter model built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), with native vision capabilities and a 1-million-token context window. It is the world's first open 3T-class model, designed for frontier intelligence across long-horizon coding, knowledge work, and reasoning. ### Key Features - **New Architecture**: Kimi K3 is built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), and scales up MoE sparsity with a Stable LatentMoE framework that activates 16 out of 896 experts — yielding an approximate 2.5× improvement in overall scaling efficiency over Kimi K2. - **Long-Horizon Coding**: Operating with minimal human oversight, Kimi K3 sustains long engineering sessions, navigates massive repositories, and orchestrates terminal tools — from GPU kernel optimization and compiler development to vision-in-the-loop game dev, CAD, and even chip design. - **Agentic Knowledge Work**: Kimi K3 advances end-to-end knowledge work, producing deep research with interactive visualizations, widgets and dashboards, and motion design and video editing, powered by its native multimodal architecture. - **Native Multimodality & Long Context**: Kimi K3 understands text, images, and video within the same model, and supports a 1-million-token context window. - **Open Frontier Weights**: We release the full Kimi K3 model weights under the Kimi K3 License, making frontier intelligence openly available for research, deployment, and further innovation. ## 2. Model Summary
Architecture Mixture-of-Experts (MoE)
Total Parameters 2.8T
Activated Parameters 104B
Number of Layers 93
Number of Dense Layers 1
Attention-Layer Composition 69 KDA + 24 Gated MLA
Attention Hidden Dimension 7168
Number of Attention Heads 96
Latent MoE Dimension 3584
MoE Hidden Dimension (per Expert) 3072
Number of Experts 896
Selected Experts per Token 16
Number of Shared Experts 2
Vocabulary Size 160K
Context Length 1048576
Attention Mechanism KDA & Gated MLA
Activation Function SiTU-GLU
Vision Encoder MoonViT-V2
Parameters of Vision Encoder 401M
Quantization MXFP4 weights / MXFP8 activations
(quantization-aware training)
Modality Text, Image
## 3. Evaluation Results
Benchmark Kimi K3
(max)		 Claude Fable 5
(max, w/ fallback)		 GPT-5.6 Sol
(max)		 Claude Opus 4.8
(max)		 GPT-5.5
(xhigh)		 GLM-5.2
(max)
Reasoning & Knowledge
GPQA Diamond 93.5 92.6 94.1 91.0 93.5 91.2
CritPt 23.4 28.6 32.3 20.9 27.1 20.9
AA-LCR 74.7 70.0 73.7 67.7 74.3 71.3
HLE-Full 43.5 / 56.0 53.3 / 63.0 44.5 / 58.0 49.8 / 57.9 41.4 / 52.2 —
Coding
DeepSWE 67.5 70.0 73.0 59.0 67.0 46.2
ProgramBench 77.8 76.8 77.6 71.9 70.8 63.7
Terminal-Bench 2.1 88.3 88.0 88.8 84.6 83.4 82.7
FrontierSWE 81.2 86.6 71.3 66.7 64.9 67.3
SWE-Marathon 42.0 35.0 39.0 40.0 14.0 13.0
PostTrainBench 36.6 41.4 34.6 34.1 28.4 34.3
MLS-Bench-Lite 48.3 49.9 46.2 42.8 35.5 40.4
SciCode 58.7 60.2 56.1 53.5 56.1 50.5
Kimi Code Bench 2.0 72.9 76.9 64.8 71.7 69.0 64.2
Agentic
BrowseComp 91.2 88.0 90.4 84.3 84.4 —
DeepSearchQA (F1) 95.0 94.2 — 93.1 — —
ResearchRubrics 76.2 — 73.8 73.5 64.0 71.1
GDPval-AA v2 (Elo) 1686 1747 1736 1593 1491 1510
Toolathlon-Verified 76.5 77.9 74.9 76.2 73.5 59.9
MCPMark-Verified 94.5 87.4 92.9 76.4 92.9 —
MCP-Atlas 84.2 84.7 83.6 83.6 82.8 82.6
AutomationBench 30.8 29.1 29.7 27.2 22.7 12.9
JobBench 54.3 57.4 45.4 48.4 38.3 43.4
AA-Briefcase (Elo) 1548 1583 1495 1354 1158 1260
Agents' Last Exam 28.3 25.7† 29.6 27.0 26.6 20.4
APEX-Agents 41.0 43.3 39.9 39.4 38.5 35.6
OfficeQA Pro 63.3 69.9 63.2 63.9 60.9 41.4
SpreadsheetBench 2 34.8 34.7 32.4 31.6 29.1 28.1
OSWorld-Verified 84.8 85.0 83.0 83.4 79.0 —
OSWorld 2.0 58.3 66.1 62.6 55.7 49.5 —
SaaS-Bench 60.1 — 61.4 56.1 43.8 —
τ³-Banking 33.4 26.8 33.0 27.6 31.3 26.8
Harvey Lab-AA 94.6 93.6 87.2 91.1 86.3 91.0
CorpFin v2 71.6 71.8 64.4 66.7 68.4 66.1
Finance Agent v2 54.4 56.3 53.8 53.9 51.8 49.7
Legal Research Bench 44.2 49.5 48.1 43.8 40.4 31.3
Vision
WorldVQA ForceAnswer 51.0 56.7 41.8 39.1 38.5 —
OmniDocBench 91.1 89.8 85.8 87.9 89.4 —
PerceptionBench 58.5 57.2 59.7 47.2 55.8 —
Video-MME (w. sub) 90.0 — 89.5 86.0 89.3 —
MMVU 82.1 — 81.2 79.2 81.7 —
BabyVision w/ python 85.7 90.5 88.9 81.2 83.6 —
MMMU-Pro 81.6 / 83.4 81.2 / 86.5 83.0 / 84.6 78.9 / 82.7 81.2 / 83.2 —
CharXiv (RQ) 84.8 / 91.3 88.9 / 93.5 84.6 / 89.1 80.5 / 89.9 84.1 / 89.0 —
MathVision 94.3 / 97.8 94.8 / 98.6 95.8 / 97.8 86.7 / 97.1 92.2 / 96.8 —
ZeroBench (pass@5) 23.0 / 41.0 23.0 / 46.0 17.0 / 35.0 17.0 / 34.0 22.0 / 41.0 —
Footnotes All Kimi K3 results are obtained with reasoning effort set to 'max' and temperature = 1.0. For single-step tasks, such as GPQA Diamond, HLE-Full, and vision benchmarks without tools, we set top-p = 0.95; for agentic tasks, we set top-p = 1.0. For HLE-Full, MMMU-Pro, CharXiv (RQ), MathVision, and ZeroBench, each cell reports the scores without and with tool augmentation (general tools for HLE-Full, Python for the vision benchmarks), in that order. 1. **Reasoning & knowledge benchmarks** - **CritPt and AA-LCR.** Scores are cited from [Artificial Analysis](https://artificialanalysis.ai/) as of July 23, 2026. 2. **Coding benchmarks** - **DeepSWE.** Kimi K3 is evaluated with the Kimi Code harness. The GLM-5.2 score is taken from the [GLM-5.2 release blog](https://z.ai/blog/glm-5.2); all remaining scores are from the official [DeepSWE leaderboard](https://deepswe.datacurve.ai/), under which Kimi K3 attains 67.3 with the mini-SWE-agent harness. We report the DeepSWE v1.1 tasks. - **Terminal-Bench 2.1.** Kimi K3 is evaluated with the Kimi Code harness. For all other models, we report the best score across harnesses: GLM-5.2 with Claude Code ([GLM-5.2 release blog](https://z.ai/blog/glm-5.2)); Claude Opus 4.8 and Claude Fable 5 with Terminus 2 ([Artificial Analysis](https://artificialanalysis.ai/evaluations/terminalbench-v2-1)); GPT-5.5 and GPT-5.6 Sol with Codex ([OpenAI](https://openai.com/index/previewing-gpt-5-6-sol/)). - **ProgramBench.** Kimi K3 is evaluated with the Kimi Code harness. The GLM-5.2 score is from the [GLM-5.2 release blog](https://z.ai/blog/glm-5.2); all other scores are from [Vals AI](https://www.vals.ai/benchmarks/programbench). - **SWE-Marathon.** Kimi K3, Claude Opus 4.8, and Claude Fable 5 are evaluated with the Claude Code harness; GPT-5.6 Sol is evaluated with the Codex harness. The GLM-5.2 score is from the [GLM-5.2 release blog](https://z.ai/blog/glm-5.2). Our evaluation is based on an H20-calibrated branch of the [official tasks](https://www.swe-marathon.org/) as of July 9, 2026, prior to the final v1.1 release: the Docker images, performance gates, and reference oracles for the GPU tasks have been recalibrated for H20, while the correctness and anti-cheat validators remain unchanged. Additionally, Claude Fable 5 hit fallbacks on 35% of the tasks in our evaluation, which may have negatively impacted its measured performance. - **FrontierSWE.** Kimi K3 is evaluated with the Kimi Code harness and GPT-5.6 Sol with the Codex harness; all other results are from [FrontierSWE](https://www.frontierswe.com/). Dominance scores are recomputed from the raw scores using the official evaluation script and are current as of July 16, 2026. - **PostTrainBench.** Scores for GLM-5.2, GPT-5.5, and Claude Opus 4.8 are adopted from the official [PostTrainBench](https://posttrainbench.com/) results. Kimi K3, Claude Fable 5, and GPT-5.6 Sol are evaluated with the official Harbor implementation at maximum reasoning effort, averaged over three runs on H20 GPUs (instead of H100 in the official setting) — Kimi K3 and Claude Fable 5 with the Claude Code harness, and GPT-5.6 Sol with the Codex harness. - **MLS-Bench-Lite.** Kimi K3 is evaluated with the Kimi Code harness; GLM-5.2 and the Claude models with the Claude Code harness; GPT-5.5 and GPT-5.6 Sol with the Codex harness. - **SciCode.** Scores are cited from [Artificial Analysis](https://artificialanalysis.ai/) as of July 23, 2026. - **Kimi Code Bench 2.0 (in-house).** Kimi K3 is evaluated with the Kimi Code harness (it attains 73.7 with the Claude Code harness); GLM-5.2, Claude Opus 4.8, and Claude Fable 5 with the Claude Code harness; GPT-5.5 and GPT-5.6 Sol with the Codex harness. All models are evaluated at maximum reasoning effort, except GPT-5.5, which uses the "xhigh" setting. As the benchmark includes cybersecurity and safety-related tasks, we also disclose the fraction of refused or fallback tasks: Claude Fable 5 hit 13 fallbacks and 1 refusal out of 80 tasks; 10 refusals out of 80 tasks entered GPT-5.6 Sol's cyber guard; GPT-5.5 had 3 refusals out of 80 tasks. 3. **Agentic benchmarks** - **OfficeQA Pro.** Each test case provides the agent with the entire PDF corpus, with all PDFs rendered as images and no machine-readable text available. - **OfficeQA Pro and SpreadsheetBench 2.** Kimi K3, GLM-5.2, Claude Opus 4.8, and Claude Fable 5 are evaluated with the Claude Code harness; GPT-5.5 and GPT-5.6 Sol are evaluated with the Codex harness. - **MCP-Atlas.** All models are evaluated on the 500-task public subset with a 100-turn limit, using Gemini 3.1 Pro as the judge. - **AutomationBench.** All models are evaluated on the 600-task public subset, following the official GitHub setup in all other respects. - **BrowseComp.** We adopt a context-compaction strategy triggered at 300K tokens. When evaluated with the full 1M-token context window and no context management, Kimi K3 achieves a score of 90.4. The results of Claude Fable 5, Claude Opus 4.8, GPT-5.6 Sol, and GPT-5.5 are cited from [Anthropic](https://www.anthropic.com/news/claude-fable-5-mythos-5) and [OpenAI](https://openai.com/index/gpt-5-6/). - **GDPval-AA v2, AA-Briefcase, τ³-Banking, Harvey Lab-AA, and APEX-Agents.** Scores are cited from [Artificial Analysis](https://artificialanalysis.ai/) and the [APEX-Agents leaderboard](https://www.mercor.com/apex/apex-agents-leaderboard/) as of July 23, 2026. For Harvey Lab-AA, we report the criterion pass rate. - **CorpFin v2, Finance Agent v2, and Legal Research Bench.** Scores are cited from [Vals AI](https://www.vals.ai/). - **Agents' Last Exam.** Scores are cited from the [official leaderboard](https://agents-last-exam.org/leaderboard) as of July 23, 2026; we report the leaderboard's primary pass-rate metric. On the leaderboard, each model is paired with a specific harness: Kimi K3 with Kimi Code; GPT-5.6 Sol and GPT-5.5 with Codex; Claude Fable 5, Claude Opus 4.8, and GLM-5.2 with Claude Code. † The Claude Fable 5 entry runs at xhigh effort with 40% of tasks annotated as downgraded. 4. **Multimodal benchmarks** - Except for ZeroBench, which follows the official setting and is run five times, all multimodal scores are averaged over three runs. MMMU-Pro is evaluated following the official protocol, preserving the original input order and prepending images to the text input. - **PerceptionBench** is an in-house benchmark that focuses on atomic visual perception capabilities.
## 4. Native MXFP4 Quantization Kimi K3 applies quantization-aware training from the SFT stage onward, using MXFP4 weights with MXFP8 activations for broad hardware compatibility. ## 5. Deployment > [!Note] > You can access Kimi K3's API on https://platform.kimi.ai by selecting `kimi-k3`, and we provide OpenAI/Anthropic-compatible API for you. Currently, Kimi K3 is recommended to run on the following inference engines: - [vLLM](https://github.com/vllm-project/vllm) — see [recipes](https://recipes.vllm.ai/moonshotai/Kimi-K3) - [SGLang](https://github.com/sgl-project/sglang) — see [cookbook](https://docs.sglang.io/cookbook/autoregressive/Moonshotai/Kimi-K3) - [TokenSpeed](https://github.com/lightseekorg/tokenspeed) — see [recipes](https://lightseek.org/tokenspeed/recipes/models#kimi-k3) --- ## 6. Model Usage Kimi K3 always has thinking enabled, and will return `reasoning_content`. Thinking effort is configured with the top-level `reasoning_effort` request field, which supports `"low"`, `"high"`, and `"max"` (default `"max"`). Kimi K3 was trained in the preserved thinking history mode. For multi-turn conversations and tool calls, Kimi K3 requires the complete assistant message returned by the API to be passed back to `messages` as-is — including `reasoning_content` and `tool_calls`, not just `content`: ```python import openai def chat_with_preserved_thinking(client: openai.OpenAI, model_name: str): messages = [ { "role": "user", "content": "Tell me three random numbers." }, { "role": "assistant", "reasoning_content": "I'll start by listing five numbers: 473, 921, 235, 215, 222, and I'll tell you the first three.", "content": "473, 921, 235" }, { "role": "user", "content": "What are the other two numbers you have in mind?" } ] response = client.chat.completions.create( model=model_name, messages=messages, stream=False, max_tokens=4096, reasoning_effort="max", ) # the assistant should mention 215 and 222 that appear in the prior reasoning content print(f"response: {response.choices[0].message.reasoning}") return response.choices[0].message.content ``` For full guides and examples (vision input, structured output, partial mode, tool choice, dynamic tool loading, context caching), see the [Kimi K3 Quickstart](https://platform.kimi.ai/docs/guide/kimi-k3-quickstart) and [Thinking Effort](https://platform.kimi.ai/docs/guide/use-thinking-effort). ### Coding Agent Framework Kimi K3 works best with [Kimi Code CLI](https://www.kimi.com/code) as its agent framework. We warmly invite you to give it a try — run Kimi Code in your terminal and select Kimi K3 using the `/model` command. We hope you enjoy building with Kimi K3, and we would love to hear your feedback! --- ## 7. License Both the code repository and the model weights are released under the [Kimi K3 License](https://huggingface.co/moonshotai/Kimi-K3/blob/main/LICENSE). --- ## 8. Contact Us If you have any questions, please reach out at [support@moonshot.ai](mailto:support@moonshot.ai).