Moonshot AI has released the Kimi K3 weights for free on Hugging Face, after earlier publishing a blog post and API documentation for the model. The Chinese AI company’s release means organizations with modern AI GPU racks can run the model themselves and, under Moonshot’s license, charge for access with relatively few stated restrictions.
The move puts an open-weight system in the same conversation as proprietary frontier models from OpenAI and Anthropic. According to Moonshot’s own benchmarks and technical report, Kimi K3 beats earlier generations of GPT and Claude models and comes close to OpenAI’s GPT-5.6 Sol and Anthropic’s Claude Fable. Those are vendor-provided results, so they should be treated as claims until more independent testing lands.
What are Kimi K3 weights?
Model weights are the numerical parameters learned during training. Releasing them lets other operators run inference on their own hardware instead of sending every request through Moonshot’s hosted API.
Open-weight access is narrower than open source. Moonshot has released the parameters needed to run Kimi K3, but it has not disclosed the full training process or dataset, which remain controlled by the company.
How much does Moonshot say Kimi K3 costs to run?
Moonshot says its inference cost comparisons use internal measurements for Kimi K3 and publicly available token pricing for rival models. For standard, non-cached input, Moonshot lists Kimi K3 at $3 per million tokens. The same comparison puts Claude Fable at $10 per million tokens and GPT-5.6 Sol at $5 per million tokens.
The company also says Kimi K3’s cache can change that math for workloads that reuse context. Moonshot claims a 90% cache hit ratio for coding tasks, which would bring cached input pricing down to $0.30 per million tokens for use cases that hit the cache heavily. Moonshot says output token comparisons follow a similar pattern, though the figures cited for input are the clearest public numbers in its materials.
Why might Kimi K3 be cheaper to run?
Moonshot’s technical report points to several inference optimizations. Kimi K3 is a 2.8 trillion-parameter model, but only 104.2 billion parameters are active at a time. It uses a sparse mixture-of-experts design, activating 16 experts out of 896 for a given step.
The model also uses MXFP4 for weights and MXFP8 for input activations, lower-precision formats that can reduce memory pressure. Moonshot says Kimi K3 avoids a conventional ever-growing key-value cache by using a fixed-size state handler called Kimi Delta Attention, a design intended to cut VRAM use and execution overhead during inference.
Moonshot’s report mentions Nvidia’s H20 GPU for some coding tests. That chip does not natively support MX floating-point types, unlike Nvidia’s export-controlled Blackwell B-series chips. That detail leaves room for two readings: Kimi K3 may already run efficiently on less capable AI hardware, and it may run better on silicon with native MXFP support. Moonshot has not provided confirmed figures for that second scenario.
For OpenAI and Anthropic, the uncomfortable part is distribution as much as performance. If Moonshot’s claims hold up, Kimi K3 gives third-party providers a model they can host themselves, price aggressively, and sell against closed systems. The benchmarks still need outside scrutiny, but the weights are now available for others to test instead of merely reading the launch claims.
This story draws on original reporting from Tom's Hardware.
