Foreword¶

On July 27, 2026, Moonshot AI simultaneously released the full model weights of Kimi K3 on Hugging Face and GitHub, and uploaded the arXiv technical report (2607.24653). This was the largest open-source 3T-scale model at the time: with a total parameter count of 2.8T, 104B of active parameters per forward pass, a 1-million-token context window, and natively supporting multimodal inputs.

For developers, the significance of this release is not that “another large model has been added”, but that the open-source community can for the first time obtain the full model weights to conduct self-hosted experiments in scenarios such as Agents, long-range encoding, and million-token context. The GitHub repository moonshotai/kimi-k3 quickly gained thousands of stars after its release, and the related discussion thread on Hacker News accumulated more than 870 points, indicating high community attention.

This article sorts out the architectural highlights, capability boundaries, and acquisition methods of Kimi K3 based on the official README, arXiv abstract, and public evaluation data.

Release Date and Acquisition Channels¶

Based on the GitHub repository creation time and arXiv submission records, Kimi K3 was officially released on July 27, 2026:
- Model Weights: Hugging Face — moonshotai/Kimi-K3
- Code and Documentation: GitHub — moonshotai/kimi-k3
- Technical Report: arXiv:2607.24653

Moonshot positions it as “Open Frontier Intelligence”, an open-weight release for cutting-edge intelligent tasks. The weights are stored in MXFP4 quantization (with activations in MXFP8), and the official introduced quantization-aware training during the SFT phase to balance hardware compatibility during deployment.

Model Architecture: KDA, Stable LatentMoE and Hybrid Attention¶

The core architectural innovations of Kimi K3 can be summarized into three layers:

1. Kimi Delta Attention (KDA) and Attention Residuals (AttnRes)
KDA is an improved attention mechanism, paired with AttnRes to improve the efficiency of information transmission across sequence length and model depth. K3 has a total of 93 Transformer layers, among which 69 layers use KDA and 24 layers use Gated MLA (Multi-head Latent Attention), forming a hybrid attention structure. The hidden size is 7168, and the number of attention heads is 96.

2. Stable LatentMoE
The routing stability and training efficiency of traditional MoE become bottlenecks when the number of experts is expanded. K3 adopts the Stable LatentMoE framework: there are 896 routing experts in total, 16 of which are activated per token, and there are 2 shared experts. The latent MoE dimension is 3584, and the hidden size of a single expert is 3072. According to the arXiv abstract, compared with Kimi K2, the overall scaling efficiency is improved by about 2.5 times (this is an official claim by Moonshot, and independent reproduction remains to be verified by the community).

3. Native Multimodality
The visual encoder is MoonViT-V2 (401M parameters), and the model natively supports text and image inputs, with a context length of 1,048,576 tokens (about 1 million). The vocabulary size is 160K, and the activation function is SiTU-GLU.

Core Parameter Overview¶

Evaluation Performance: Agent and Coding as the Main Battlefields¶

Moonshot published a large number of benchmark results in the README (with the reasoning effort set to max). It should be emphasized that Agent benchmarks highly depend on the harness (agent framework), and scores under different frameworks cannot be directly compared horizontally. The following data are all from the official README, and third-party independent verification results may vary slightly.

Coding and Terminal Tasks¶

• Terminal-Bench 2.1: 88.3 (Kimi Code harness). GPT-5.6 Sol scored 88.8 and Claude Fable 5 scored 88.0 on the same benchmark. Independent evaluation organization Artificial Analysis reported that K3 scored about 85% on its own harness, with a harness difference from the official 88.3. Pay attention to the comparison conditions when reading the leaderboard.
• DeepSWE: 67.5 (Kimi Code); 67.3 when using the mini-SWE-agent harness, with a very small gap.
• FrontierSWE: 81.2 (dominance score)
• ProgramBench: 77.8
• SWE-Marathon: 42.0 (multi-hour full project engineering tasks)

Reasoning and Knowledge¶

• GPQA Diamond: 93.5
• AA-LCR (long-context reasoning): 74.7

Agent Tasks¶

• BrowseComp: 91.2 (context compaction strategy triggered at 300K tokens)
• MCPMark-Verified: 94.5

The conclusion of the arXiv abstract states that K3 reaches the cutting-edge level in long-range encoding, Agent, knowledge, reasoning, and vision tasks; overall, it is still slightly inferior to Claude Fable 5 and GPT-5.6 Sol, but consistently outperforms other open-source and some closed-source models in its evaluation suite. This statement is relatively restrained, which is basically consistent with the community’s perception of “close to Fable and open-source SOTA”.

Post-Training and Infrastructure¶

The technical report focuses on three types of post-training and system engineering:
1. Multi-domain Reinforcement Learning: Covers general, Agent, coding and other scenarios, supports multiple reasoning effort levels, and emphasizes combinatorial generalization and long-range execution capabilities.
2. Million-token Agentic RL: Supports persistent rollout and sandbox state retention, suitable for long tool call chain training.
3. Algorithm-System Co-design: Load balancing for parallel training of KDA and experts, efficient memory management, and deployment-side innovations.

These engineering details explain why K3 can maintain usability in million-context and complex Agent scenarios, rather than just relying on stacking parameters.

How to Deploy and Invoke¶

API Invocation¶

Moonshot provides an OpenAI/Anthropic-compatible API at platform.kimi.ai, just select kimi-k3 as the model name. K3 enables thinking mode by default, and the response contains reasoning_content; for multi-turn conversations, you need to pass back the complete assistant message (including reasoning and tool_calls) as-is.

import openai

client = openai.OpenAI(
    base_url="https://api.moonshot.ai/v1",
    api_key="YOUR_API_KEY",
)

response = client.chat.completions.create(
    model="kimi-k3",
    messages=[{"role": "user", "content": "解释 Stable LatentMoE 的路由机制"}],
    max_tokens=4096,
    reasoning_effort="max",  # Optional: low / high / max
)
print(response.choices[0].message.content)

Local Inference¶

The official recommends the following inference engines and provides corresponding recipes:
- vLLM — Kimi-K3 recipes
- SGLang — cookbook
- TokenSpeed

Agent Framework¶

The official recommends using Kimi Code CLI as the Agent framework, select Kimi K3 via the /model command in the terminal to achieve the best performance in scenarios such as Terminal-Bench.

Self-hosting the full model weights has extremely high requirements for GPU clusters (the weight volume after MXFP4 quantization is still in the hundreds of GB range). Most teams will first verify the effect through the API or cloud hosting services, and then evaluate whether it is worth investing in local deployment.

License and the Boundary of “Open Source”¶

Both the weights and code are released under the Kimi K3 License, which allows research, modification, and distribution, but has additional clauses for large-scale Model-as-a-Service commercial use and ultra-large-scale commercial products. Therefore, the community often refers to it as open-weight (open weights), rather than fully open source in the traditional sense (such as Apache 2.0). It is recommended to read the full text of the LICENSE before use to confirm whether it meets your scenario.

Summary¶

Kimi K3 is currently one of the cutting-edge models in the open-source ecosystem with the largest parameter scale, longest context, and explicitly targeting Agent and long-range coding. The figures of 2.8T MoE, 104B active parameters, 16 routed experts out of 896, 1-million-token context, and native multimodality jointly define a new benchmark for open-source large models in the second half of 2026.

For ordinary developers, it is recommended to experience it through the API or Kimi Code first; for research institutions and teams with sufficient computing power, the full model weights can be downloaded on Hugging Face, and used with vLLM/SGLang for secondary development and evaluation reproduction. Benchmark scores must be interpreted in combination with the harness and evaluation conditions, and should not be simply equated to “model bare capability ranking”.

Whether you plan to self-host or not, the release of K3 marks another important move by a Chinese team in the open-source cutting-edge model track. The competition of large models in the Agent era is moving from “API available” to “weights available, modifiable, and verifiable”.