Kimi K3 open weights are live and 1.4 terabytes
At midnight UTC, Moonshot AI flipped the switch. The full weights of Kimi K3, a 2.8-trillion-parameter model, are now downloadable by anyone. It is the largest open-weight release in history.
The file size is 1.4 terabytes in MXFP4 quantization. That is not a typo. Downloading the model is the easy part. Running it requires roughly 18 H100 80GB GPUs, which puts the practical entry point at around $50 per hour on reserved cloud capacity. Moonshot recommends 64 or more accelerators for production use. Free to download does not mean free to run.
The model itself is a Mixture of Experts architecture with 896 experts, activating 16 per token. Moonshot introduced two architectural changes: Kimi Delta Attention (KDA) for better information flow across long sequences, and Attention Residuals (AttnRes) for improved depth. They claim a 2.5x scaling efficiency improvement over Kimi K2. The context window is one million tokens, enough to feed entire codebases into a single prompt.
Performance-wise, the honest framing is this: K3 is a specialist, not a category-killer. Independent benchmarks place it below Claude Fable 5 and GPT-5.6 Sol on overall capability. But it topped a specific front-end coding leaderboard in blind arena evaluations, and it consistently beats other open models. Arena evaluations showed developers picking K3 ahead of US frontier models on coding tasks. On GPU kernel optimization, Moonshot’s own benchmarks show K3 competitive with Fable 5 and ahead of GPT-5.6 Sol. The model also built MiniTriton, a working Triton-like GPU compiler from scratch, including IR passes and PTX codegen.
K3’s security capabilities are not theoretical. Last week, a researcher using Kimi K3 agents found 19 zero-day vulnerabilities in Redis in 90 minutes and built a working RCE exploit in 27 minutes. Redis shipped seven emergency security releases. The vulnerabilities require authentication, so the attack surface is narrower than the headline suggests, but the speed of discovery is the signal. This is the same model now available for anyone to download and run.
The business context explains why Moonshot is giving away a model that cost hundreds of millions to train. Since K3 launched on the API two weeks ago, Moonshot’s daily revenue grew by at least 6x. The company hit $300 million in annual recurring revenue in June, up from $200 million in April. They are seeking a new funding round at a $50 billion valuation, with a potential Hong Kong IPO as soon as this year. Open weights are a distribution strategy, not charity.
The weight release also intensifies the political fight with Washington. White House science director Michael Kratsios accused Moonshot last week of training K3 using banned Nvidia chips and conducting large-scale distillation against US models. Moonshot has not responded. The weights are now public regardless, hosted on Hugging Face, beyond the reach of export controls.
The immediate beneficiaries are inference providers, not individual developers. Companies like Together AI, Fireworks, and similar specialized hosts can serve K3 through their existing multi-GPU infrastructure. None of the three major hyperscalers (AWS Bedrock, Azure Foundry, Google Vertex AI) have integrated Kimi K3 or any comparable Chinese open-weight model. The community will produce further-quantized versions that fit smaller hardware, but that takes time.
The real advantage of self-hosting K3 is data control. Running the model on your own hardware means no API calls to a Beijing-based company, no questions about data provenance, and no dependency on a provider that could be sanctioned. For organizations that need frontier-level coding capability and cannot or will not send data to US or Chinese API providers, this is the first open model that gets close enough to matter.
Moonshot also published a technical report alongside the weights covering architecture, training methodology, and evaluations. The model supports native vision capabilities, meaning it can process images and screenshots alongside text, which is how it handles tasks like game development and frontend coding with visual feedback loops.
The open-weight tier now has a new ceiling. Before K3, the largest open models hovered around 700 billion parameters. Moonshot just quadrupled that. Whether the next move comes from DeepSeek, Qwen, or a Western lab responding to the pressure, the gap between open and closed is narrowing faster than expected.