Moonshot AI just released the full model weights for Kimi K3, a 2.8 trillion parameter AI model that sits right behind GPT-5.6 and Claude Fable 5 on benchmarks. It beat Claude in frontend coding. It builds GPU compilers from scratch. And now you can download the whole thing.
What just happened
On July 27, Moonshot AI published the complete Kimi K3 weights on Hugging Face. That is 1.56 TB of model files, 96 safetensors shards, available for anyone to download. This is the first open 3T-class model ever released, and it is not a toy. Kimi K3 scores 57 on the Artificial Analysis Intelligence Index, ahead of every other open-weight model and trailing only three closed ones: GPT-5.6 Sol, Claude Fable 5, and Claude Opus 5.
The model itself launched on July 16, but the open-weight release is what changes the game. Eleven days after going live in the Kimi app, the weights are public. The technical report shipped alongside them, not months later.
What Kimi K3 can actually do
The benchmarks are impressive, but the case studies are where K3 gets interesting.
Frontend coding that beats Claude. Kimi K3 topped a frontend coding benchmark ahead of Claude, which is not a small thing. If you have ever wrestled with getting an AI model to produce clean, working frontend code, this matters.
Game development with vision in the loop. K3 can take a concept, image, or video and turn it into a fully playable interactive experience. It does this by iterating between code and live screenshots, seeing what it built and refining it in real time. That is a workflow that used to take a designer and a developer working together.
GPU compiler development from scratch. In testing, K3 built MiniTriton, a compact Triton-like GPU compiler with its own IR layer, optimization passes, and PTX code generation pipeline. It beat Triton on certain workloads. A model building a compiler that rivals an extensively optimised human-built one is a serious flex.
Chip design in 48 hours. K3 designed a chip to serve a nano model built on its own architecture. In a single autonomous 48-hour run, it built, optimised, and verified the chip using open-source EDA tools. The chip closes timing at 100 MHz and sustains over 8,700 tokens per second in simulation. A chip designed by a model, for a model.
Research automation. K3 reproduced the I-Love-Q universal relations in computational astrophysics in about two hours. That work typically takes one to two weeks for an experienced researcher. It reviewed 20+ papers, implemented the numerical pipeline, evaluated 300+ equations of state, identified inconsistencies in published formulas, and produced an interactive HTML dashboard.
What is inside
Kimi K3 uses a Mixture-of-Experts architecture with 2.8 trillion total parameters, but only activates 104 billion per token. That sparsity ratio is how Moonshot gets frontier performance without frontier compute costs on every inference call.
Key specs:
- 2.8T total parameters, 104B active per token
- 1 million token context window
- 93 layers: 69 Kimi Delta Attention, 24 Gated MLA
- 896 experts, 16 selected per token
- Native vision: accepts text, image, and video input
- Reasoning effort control: low, high, or max (default max)
- Weights ship in MXFP4 with MXFP8 activations
The architecture includes two novel components: Kimi Delta Attention (KDA), which enables up to 6.3x faster decoding in million-token contexts, and Attention Residuals (AttnRes), which improves information flow across model depth. Together, they deliver roughly a 2.5x improvement in scaling efficiency over Kimi K2.
What the licence actually says
This is where you need to pay attention. The Kimi K3 License is not MIT.
You can use, modify, distribute, fine-tune, and sell the model freely, with two conditions. First, if your business generates more than $20 million in revenue over any 12-month period from reselling model access, you need a separate agreement with Moonshot. Second, if your product has more than 100 million monthly active users or more than $20 million in monthly revenue, you have to display “Kimi K3” prominently in the interface.
Both conditions are waived for purely internal use. For most developers, startups, and research teams, the practical impact is zero. For hyperscalers, it is a negotiation. This is less permissive than GLM’s MIT licence but more permissive than every closed model.
What this unlocks
For developers building products. You now have a frontier-class model you can run on your own infrastructure, fine-tune on your own data, and deploy without sending every request to someone else’s API. The $3 per million input tokens on the official API is the same headline rate as Claude Sonnet 5, but self-hosting lets you avoid per-token costs entirely if you have the hardware.
For the open-source ecosystem. This is the first time a model this capable has been fully open-weight. Previous “open” models at this scale either did not release weights or released them with heavy restrictions. K3’s weights, architecture, and training details are all public. Researchers can study how it works, not just what it outputs.
For creative and design workflows. K3’s vision capabilities mean it can take screenshots and iterate on visual work. That “vision in the loop” approach is what makes it good at game development and frontend coding. You can describe what you want, see what it produces, and it refines based on what it sees.
What to look out for
Running K3 yourself requires serious hardware. The 1.56 TB checkpoint needs a multi-node GPU cluster to serve at full context. This is not a model you download and run on a Mac Mini. For practical self-hosting on smaller setups, you would need quantised versions or smaller Kimi models, which are not part of this release.
The 51% hallucination rate reported in some benchmarks is worth watching. K3 is powerful, but like every large model, it can confidently generate incorrect information. The reasoning modes help, but verification is still essential.
The licence’s revenue gates mean this is genuinely open for individuals and most companies, but starts to differ from truly permissive licences once you scale past $20 million in revenue from model-as-a-service. Read the terms before building a commercial product on top of it.
How to get started
The weights are on Hugging Face at moonshotai/Kimi-K3. The API is available at platform.kimi.ai with pricing at $3 per million input tokens and $15 per million output tokens, with cached input at $0.30. Moonshot recommends vLLM, SGLang, or TokenSpeed as inference runtimes for self-hosting.
Kimi K3 is available now on kimi.com, Kimi Work, Kimi Code, and the Kimi API.
Want more like this? Join MOKU CLUB for free. Weekly resources, early access to new guides, and occasional templates you can actually use. Join below.
