The Chinese startup Moonshot AI has released the full weights of Kimi K3, a model with 2.8 trillion parameters and a context window of 1 million tokens. The license, however, limits its commercial use for companies with revenues exceeding 20 million dollars.
Moonshot AI has launched the full weights of Kimi K3, an artificial intelligence model with 2.8 trillion parameters and a context window of 1 million tokens. This is the first model of the so-called '3T class' to be offered with open weights to the public, although with significant commercial usage restrictions.
The model, built on a sparse Mixture-of-Experts (MoE) architecture, is designed for programming tasks, autonomous agents, complex reasoning, and long-horizon knowledge work. According to independent evaluations, Kimi K3 rivals or surpasses proprietary models such as OpenAI's GPT-5.5 and Anthropic's Claude Opus 4.8 in programming and reasoning benchmarks.
The Kimi K3 license is described as a modified variant of MIT, but includes clauses that limit its corporate adoption. Companies with annual revenues exceeding 20 million dollars need to negotiate a separate commercial license to use the model in production. Additionally, operators offering Kimi K3 as an inference service to third parties (Model-as-a-Service) face prohibitions or specific requirements, regardless of their size.
Visible attribution to Moonshot AI is also required in large-scale products integrating the model, similar to the requirements of licenses like Apache 2.0 or Meta's Llama License. These restrictions follow the precedent set by Kimi K2 and align with the strategy of Chinese and Western labs to use open weights as a competitive lever without fully relinquishing commercial control.
For startups and SMEs, the license allows experimentation and deployment of the model without initial barriers, as long as they do not exceed the revenue threshold. This presents an opportunity to access a cutting-edge model with full control over data and infrastructure, although founders must conduct prior legal analysis before integrating K3 into commercial products.
The technical specifications of Kimi K3 include native multimodality with integrated visual understanding, and an API available from launch with pricing of $3 per million input tokens, $0.30 with cache, and $15 per million output tokens. The 1 million token context window allows for processing extensive documents, complete codebases, or prolonged agent sessions.
Compared to other open-weight models, Kimi K3 stands out for its extreme scale and massive context. While Meta's Llama 3 reaches up to 405 billion parameters and 256,000 context tokens, and Mistral Large does not disclose its parameter count, Kimi K3 offers 2.8 trillion parameters in MoE architecture and 1 million context tokens. Alibaba's Qwen 2.5, on the other hand, reaches up to 72 billion parameters.
The decision by Moonshot AI reflects a growing trend in the open-weight AI ecosystem: to democratize technical access without relinquishing competitive advantage in high-value business segments. Chinese labs, in particular, are using this strategy to rapidly distribute their models and encourage community adoption while preserving monetization opportunities.
For companies evaluating AI options for in-house deployment, Kimi K3 represents an attractive alternative, but they must carefully consider the license restrictions. The key is to determine whether the model fits their business model and whether they are willing to negotiate a commercial license if they exceed the revenue threshold.
Moonshot AI has already made the weights and technical documentation available in its official repository. The developer and startup community is expected to begin exploring its capabilities, especially in areas such as knowledge-intensive process automation and the creation of autonomous agents.
You may also like
