Moonshot AI has released the full weights for Kimi K3 on Hugging Face, giving developers and researchers access to what the Chinese company describes as the world’s first open model in the three-trillion-parameter class.
The model contains 2.8 trillion total parameters, making it the largest open-weight AI system disclosed by parameter count. Yet Kimi K3 does not use all of that capacity for every word it generates. Its highly sparse Mixture-of-Experts architecture activates 16 of 896 routed experts for each token, along with two shared experts, leaving about 104 billion parameters active at a time.
That distinction is central to the design. The headline number reflects the model’s total stored capacity, while the smaller active set determines much of the computing required during inference. Moonshot says the architecture delivers about 2.5 times better scaling efficiency than Kimi K2, though the model remains exceptionally demanding to host.
Kimi K3 is built for long-running professional tasks rather than short chatbot exchanges. Moonshot is positioning it for software engineering, document analysis, deep research, tool use, interactive dashboards, visual design, and projects that may require many stages of planning and execution.
The release gives the open AI ecosystem a model that approaches the scale of the most powerful proprietary systems while allowing organizations to run its weights on infrastructure they control. It also arrives as American startups debate possible restrictions on Chinese open-weight models, which have become increasingly important to developers seeking lower costs and greater control.
Extreme Scale With Only a Fraction Active
Kimi K3 relies on three major architectural choices: Kimi Delta Attention, Attention Residuals, and a Stable LatentMoE framework. Together, they are intended to improve the way information moves through both very long sequences and the depth of the model.
Kimi Delta Attention uses a hybrid approach that combines linear attention with periodic full-attention layers. This is designed to reduce the memory burden associated with very long prompts while preserving the model’s ability to retrieve precise information from earlier parts of a conversation or document collection.
Attention Residuals changes how information passes between layers. Instead of relying on a single conventional residual stream, the model can access representations from multiple earlier layers. Moonshot argues that this helps the system preserve useful information as tasks become deeper and more complex.
The model supports a context window of 1,048,576 tokens. In practical terms, that is enough capacity to place large software repositories, extensive document sets, research material, or long interaction histories into one working context, although real-world accuracy still depends on the structure and difficulty of the task.
Kimi K3 also includes Moonshot’s MoonViT-V2 vision encoder. The company says the model can understand text, images, and video natively, allowing it to combine visual perception with coding and research. That could be useful for tasks such as rebuilding an interface from screenshots, analyzing diagrams, editing video, developing games, or using visual feedback while working with design and engineering tools.
Moonshot describes Kimi K3 as capable of sustaining long engineering sessions with limited supervision. Its examples include optimizing GPU kernels, navigating large repositories, building compilers, working on computer-aided design, and using visual feedback during front-end or game development.
The Hugging Face release includes model weights and code under the Kimi K3 License. That makes K3 open-weight, but not necessarily fully open-source under the strictest definition. Moonshot has not released every component required to reproduce the model from the beginning, such as the complete training dataset and full training process.
Even with its sparse architecture and native MXFP4 quantization, Kimi K3 is not a model most developers will run on a workstation. Moonshot recommends high-bandwidth supernode configurations with at least 64 accelerators for efficient deployment. Serving software must also manage expert routing, recurrent attention state, multimodal inputs, and the enormous memory demands of a one-million-token context.
Moonshot worked with the vLLM community, NVIDIA, AMD, and other infrastructure partners before the weight release. vLLM has prepared deployment code, Docker images, caching changes, and hardware-specific optimizations intended to make the model usable in production environments from the first day of availability.
Frontier Performance With Important Caveats
Moonshot’s benchmark results place Kimi K3 close to leading proprietary models across coding, research, and tool use. The company reports a score of 88.3 on Terminal-Bench 2.1, 77.8 on ProgramBench, 81.2 on FrontierSWE, and 91.2 on BrowseComp.
On several tests, Kimi K3 matched or exceeded results reported for models from OpenAI and Anthropic. It also performed strongly on tasks involving long-horizon software work, research, browser use, automation, and the Model Context Protocol.
Those numbers require caution. Many were produced by Moonshot using its own agent framework and maximum reasoning settings, while competing systems were sometimes tested through different harnesses. Vendor benchmark tables are useful for comparing direction and capability, but they are not a substitute for broad independent testing under identical conditions.
Moonshot itself acknowledges that Kimi K3 still trails the strongest proprietary models in overall user experience. The company specifically says Claude Fable 5 and GPT-5.6 Sol remain ahead across the full range of capabilities, even though Kimi K3 is competitive or superior on selected technical evaluations.
The model card also identifies limitations. Kimi K3 can become unstable if an agent framework fails to preserve its full reasoning history, and it may act too proactively when instructions are ambiguous. Developers deploying it in sensitive workflows will need explicit boundaries, compatible software, and careful human oversight.
Kimi K3 is available through Kimi’s consumer app, Kimi Work, Kimi Code, and the Kimi API, in addition to the downloadable weights. Moonshot prices API access at $0.30 per million cached input tokens, $3 per million uncached input tokens, and $15 per million output tokens.
The Hugging Face release lands during an intense period of competition among Chinese AI developers. Alibaba recently previewed Qwen 3.8, a 2.4 trillion-parameter multimodal system whose open weights are also promised, while Z.ai, DeepSeek, and other laboratories are competing on coding performance, context length, and price.
Open weights have become a strategic advantage in that race. They let developers inspect and customize models, deploy them within private environments, and avoid dependence on a single hosted provider. They can also encourage software and infrastructure companies to optimize around a model, increasing its influence even when the original developer charges for hosted access.
The model’s scale does not automatically guarantee better results, and total parameter count can be misleading when architectures activate only a small portion of their weights. Kimi K3’s importance lies instead in the combination of scale, sparse computation, long context, multimodality, and downloadable weights.
For most developers, the model will still be accessed through an API or a specialist inference provider rather than deployed locally. For cloud companies, research institutions, and large enterprises, however, the Hugging Face release offers something unusual: access to near-frontier capabilities without requiring every prompt, document, or code repository to pass through a proprietary American platform.
That makes Kimi K3 more than another large model launch. It is a test of whether open-weight systems can move from being lower-cost alternatives to becoming the foundation for serious enterprise and agentic workloads. The answer will depend less on its 2.8 trillion parameters than on how reliably the broader ecosystem can deploy, govern, and build around them.
Disclaimer: AIstify is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and insights on artificial intelligence, emerging technologies, automation, and related industries. NuvexMedia LLC invests in and collaborates with companies across the AI, technology, software, and digital innovation sectors. These relationships do not influence AIstify’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2026 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
