Image: Moonshot AI / Hugging Face Kimi K3's 2.8-Trillion-Parameter Weights Are Now Free to Download
Moonshot AI released the full open weights for Kimi K3 on July 26-27. At 1.4TB in MXFP4 format, it's the largest open-weight model ever published. Together AI and Modal launched day-0 hosted access for teams without Blackwell hardware.
Moonshot AI released the downloadable weights for Kimi K3 on July 26, roughly seven days after the model launched on the Kimi app and API. The weights are available on Hugging Face under the moonshotai/Kimi-K3 repository.
This is the follow-through on what Moonshot announced when they launched the model on July 16: a 2.8-trillion-parameter mixture-of-experts model that would go open-weight by July 27.
What you’re actually downloading
K3 is a sparse MoE model. At inference time, roughly 50 billion parameters activate per token, not all 2.8 trillion. Per-token compute resembles a mid-size dense model. The tradeoff is that all 2.8 trillion parameters have to stay loaded in memory regardless of how many fire on any given step.
In MXFP4 (four-bit floating-point) precision, the weights need approximately 1.4 terabytes of fast memory. At standard 16-bit precision, that expands to around 5.6 terabytes. Moonshot is distributing the MXFP4 format as the primary release.
The model runs natively on Nvidia’s Blackwell generation and AMD’s MI400 accelerators. For most teams, that means multi-accelerator server nodes rather than anything you can run locally.
Day-0 hosting
Together AI and Modal both announced hosted access timed to the weights release, for teams without the hardware to run it directly. Both platforms offer API access to K3 with the same 1-million-token context window available through the Moonshot API.
The self-hosting option has a meaningful advantage for some organizations: it removes the data routing path through Chinese infrastructure. Teams with data residency requirements or concerns about the jurisdictional reach of a Chinese API provider can run K3 themselves or through a US-based cloud.
Why this matters for open-source AI
The previous largest open-weight model was DeepSeek V4 Pro at roughly 1.6 trillion parameters. K3 roughly doubles that.
Nathan Lambert at Interconnects called it the strongest open model ever released. The performance numbers are real: K3 ranked first on the Frontend Code Arena, a blind developer voting benchmark, ahead of Claude Fable 5. It also scores 81.2 on FrontierSWE and 88.3 on Terminal-Bench 2.1.
Moonshot is candid that K3 still trails Fable 5 and GPT 5.6 Sol on overall capability. But it competes on coding and agentic benchmarks with models from Anthropic and OpenAI that are not open-weight at all.
The broader context is a geopolitical one. Xi Jinping publicly backed China’s AI ecosystem releasing frontier-capable models openly at the World AI Conference this year. K3 is the most significant example of that commitment so far.
License status
At time of publication, the full license terms had not been formally published alongside the weights. Commercial usability at scale is unconfirmed until Moonshot releases complete terms. Check the Hugging Face model card for updates.
Sources: Interconnects.ai — TECHi — Quartz — TechTimes