Kimi K3

Evaluation Methodology

General Settings

• All Kimi K3 results: reasoning_effort="max", temperature=1.0
• Single-step tasks (GPQA Diamond, HLE-Full, vision benchmarks without tools): top-p=0.95
• Agentic tasks: top-p=1.0
• For HLE-Full, MMMU-Pro, CharXiv (RQ), MathVision, ZeroBench: each cell reports scores without and with tool augmentation (general tools for HLE-Full, Python for vision benchmarks)

Reasoning & Knowledge Benchmarks

• CritPt and AA-LCR: Scores cited from Artificial Analysis as of July 23, 2026

Coding Benchmarks

• DeepSWE: Kimi K3 evaluated with Kimi Code harness (67.3 with mini-SWE-agent harness); GLM-5.2 score from GLM-5.2 release blog; others from official DeepSWE leaderboard
• Terminal-Bench 2.1: Kimi K3 with Kimi Code harness; other models report best score across harnesses
• ProgramBench: Kimi K3 with Kimi Code harness; GLM-5.2 from release blog; others from Vals AI
• SWE-Marathon: Kimi K3, Claude Opus 4.8, Claude Fable 5 with Claude Code harness; GPT-5.6 Sol with Codex harness; evaluation on H20-calibrated branch
• FrontierSWE: Kimi K3 with Kimi Code harness; GPT-5.6 Sol with Codex harness; others from FrontierSWE
• PostTrainBench: Kimi K3, Claude Fable 5 with Claude Code harness, GPT-5.6 Sol with Codex harness; averaged over 3 runs on H20 GPUs
• MLS-Bench-Lite: Kimi K3 with Kimi Code harness; Claude models with Claude Code; GPT models with Codex
• SciCode: Scores from Artificial Analysis (July 23, 2026)
• Kimi Code Bench 2.0: In-house benchmark; Kimi K3 with Kimi Code harness (73.7 with Claude Code harness)

Agentic Benchmarks

• BrowseComp: Context-compaction at 300K tokens; full 1M context scores 90.4
• OfficeQA Pro: PDFs rendered as images, no machine-readable text
• MCP-Atlas: 500-task public subset, 100-turn limit, Gemini 3.1 Pro as judge
• AutomationBench: 600-task public subset, official GitHub setup
• GDPval-AA v2, AA-Briefcase, τ³-Banking, Harvey Lab-AA, APEX-Agents: From Artificial Analysis and APEX-Agents leaderboard (July 23, 2026)
• CorpFin v2, Finance Agent v2, Legal Research Bench: From Vals AI
• Agents' Last Exam: From official leaderboard (July 23, 2026); each model paired with specific harness

Multimodal Benchmarks

• Except ZeroBench (5 runs, official setting), all multimodal scores averaged over 3 runs
• MMMU-Pro: official protocol, original input order, images prepended to text input
• PerceptionBench: In-house benchmark for atomic visual perception

Citation and Resources

Primary Sources

Organization Links

Contact

• Email: support@moonshot.ai

Quickstart Guides

Thinking Control

Always-On Thinking

Kimi K3 always has thinking enabled and returns reasoning_content in every response.

Reasoning Effort Levels

The reasoning_effort request field controls thinking depth:

Preserved Thinking History

Kimi K3 was trained in preserved thinking history mode. This means:

• For multi-turn conversations, the complete assistant message (including reasoning_content and tool_calls) must be passed back to messages as-is.
• Only passing content (without reasoning_content) will degrade performance in multi-turn settings.
• This is critical for tool calls — the thinking context must be preserved across turns.

Evaluation Settings

• All benchmark results use reasoning_effort="max" with temperature=1.0
• Single-step tasks: top-p=0.95
• Agentic tasks: top-p=1.0

Additional Resources

• Thinking Effort Guide

Training Approach

Quantization-Aware Training (QAT)

Kimi K3 applies quantization-aware training from the SFT stage onward, using:

• MXFP4 weights (4-bit microscaling float)
• MXFP8 activations (8-bit microscaling float)

This means the model was trained with the quantization format in the loop, rather than being quantized post-hoc. This preserves model quality at reduced precision.

Architecture Training Stability

The Stable LatentMoE framework enables stable training of the high-sparsity MoE architecture (896 experts, 16 activated). This framework was key to achieving the 2.5× scaling efficiency improvement over Kimi K2.

Training Data

• Languages: English, Chinese
• Training data info: Quantization-aware training from SFT stage onward using MXFP4 weights with MXFP8 activations
• Full details available in the Technical Report (PDF)

Vision and Multimodal Capabilities

Native Multimodality

Kimi K3 understands text, images, and video within the same model, powered by its native multimodal architecture.

Vision Encoder

• Encoder: MoonViT-V2
• Parameters: 401M
• Supports text, image, and video inputs

Modalities

Vision Benchmark Performance

Notes

• Except for ZeroBench (5 runs, official setting), all multimodal scores are averaged over 3 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 focusing on atomic visual perception capabilities.

Model Ecosystem

Model Tree

Quantization Compatibility

The model page lists compatibility with:

• llama.cpp
• LM Studio
• Jan
• Ollama

Community

• 171 discussions on HuggingFace
• 46 Spaces using moonshotai/Kimi-K3
• 2,787,971 downloads last month
• 11 likes
• 17,600 followers for Moonshot AI

Collection

Part of the Kimi K3 Collection — 2 items, 113 likes.

Key Highlights

Kimi K3 represents several breakthroughs as Moonshot AI's most capable model:

1. World's first open 3T-class model — 2.8T parameters with open weights
2. Novel KDA Architecture — Kimi Delta Attention + Attention Residuals, a departure from standard attention mechanisms
3. 2.5× scaling efficiency over Kimi K2 via Stable LatentMoE framework (16/896 experts activated)
4. Native multimodality — text, image, and video understanding in a single model
5. 1M-token context window — native, no extension needed
6. Long-horizon coding — operates with minimal human oversight across massive repositories
7. Agentic knowledge work — deep research with interactive visualizations, dashboards, motion design
8. Native MXFP4 quantization — quantization-aware training from SFT stage for broad hardware compatibility
9. Open frontier weights — full weights under Kimi K3 License for research and commercial deployment

Context Length

Kimi K3 supports a 1-million-token context window natively. The model can process text, images, and video within this context.

Context Management

• For BrowseComp evaluation, a context-compaction strategy is triggered at 300K tokens.
• When evaluated with the full 1M-token context window and no context management, Kimi K3 achieves a BrowseComp score of 90.4 (vs. 91.2 with compaction).

License

Both the code repository and the model weights are released under the Kimi K3 License.

• The model is open-weight with commercial use allowed.
• The license terms are available at the above URL.

Coding Agent Framework

Kimi K3 works best with Kimi Code CLI as its agent framework. Users can run Kimi Code in their terminal and select Kimi K3 using the /model command.

Capabilities

• Long-Horizon Coding: Sustains long engineering sessions with minimal human oversight
• Repository Navigation: Navigates massive codebases effectively
• Terminal Tool Orchestration: From GPU kernel optimization and compiler development to vision-in-the-loop game dev, CAD, and chip design
• Multi-tool Coordination: Coordinates terminal tools, file editors, and build systems across complex workflows

Model Usage

Thinking Mode

Kimi K3 always has thinking enabled and will return reasoning_content. Thinking effort is configured with the top-level reasoning_effort request field:

• "low" — minimal reasoning
• "high" — moderate reasoning
• "max" (default) — maximum reasoning

Preserved Thinking History Mode

Kimi K3 was trained in the preserved thinking history mode. For multi-turn conversations and tool calls, the complete assistant message returned by the API must be passed back to messages as-is — including reasoning_content and tool_calls, not just content.

Example: Multi-turn with Preserved Thinking

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 from prior reasoning content
    print(f"response: {response.choices[0].message.reasoning}")
    return response.choices[0].message.content

Additional Resources

• Kimi K3 Quickstart Guide
• Thinking Effort Guide
• Full guides cover: vision input, structured output, partial mode, tool choice, dynamic tool loading, context caching.

Deployment

API Access

Kimi K3's API is available on https://platform.kimi.ai by selecting kimi-k3. The API provides OpenAI/Anthropic-compatible endpoints.

Supported Inference Engines

Native MXFP4 Quantization

Kimi K3 applies quantization-aware training (QAT) from the SFT stage onward, using:

• MXFP4 weights — 4-bit microscaling float format for model weights
• MXFP8 activations — 8-bit microscaling float format for activations

This approach ensures broad hardware compatibility while maintaining model quality. The quantization is native (baked into the model during training), not a post-hoc compression step, which preserves the model's reasoning and coding capabilities at the reduced precision.

Tensor Types in Released Weights

• F32
• BF16
• U8

The model is distributed in Safetensors format with the native MXFP4 quantization already applied.

Evaluation Results

All Kimi K3 results obtained with reasoning effort set to 'max' and temperature = 1.0. For single-step tasks (GPQA Diamond, HLE-Full, vision benchmarks without tools), top-p = 0.95; for agentic tasks, top-p = 1.0.

Reasoning & Knowledge

Coding

Agentic

Vision

Footnotes

• CritPt and AA-LCR scores cited from Artificial Analysis (July 23, 2026).
• For HLE-Full, MMMU-Pro, CharXiv (RQ), MathVision, and ZeroBench, each cell reports scores without and with tool augmentation.
• DeepSWE: Kimi K3 evaluated with Kimi Code harness (67.3 with mini-SWE-agent harness).
• BrowseComp: Context-compaction strategy triggered at 300K tokens; full 1M-token context scores 90.4.

Architecture Summary

Key Architectural Innovations

• Kimi Delta Attention (KDA): Novel attention mechanism used in 69 of 93 layers.
• Gated MLA: Used in 24 layers, complementing KDA for a hybrid attention design.
• Attention Residuals (AttnRes): Residual connections across attention layers for improved gradient flow.
• Stable LatentMoE: Framework for stable training of high-sparsity MoE models.
• SiTU-GLU: Custom activation function combining SiLU with a Gated Linear Unit.

Model Introduction

Kimi K3 is an open-weight, native multimodal agentic model and Moonshot AI's 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: Built on Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), scaling 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: Operates with minimal human oversight, 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 chip design.
• Agentic Knowledge Work: Produces deep research with interactive visualizations, widgets and dashboards, and motion design and video editing, powered by its native multimodal architecture.
• Native Multimodality & Long Context: Understands text, images, and video within the same model; supports a 1-million-token context window.
• Open Frontier Weights: Full model weights released under the Kimi K3 License for research, deployment, and further innovation.