Moonshot AI Releases Kimi K3, an Open-Weight Multimodal Model With a 1M-Token Context Window

Moonshot AI has released Kimi K3, a native multimodal model that the company describes as open-weight and aimed at long-context coding, reasoning, and knowledge work.

A large open-weight multimodal release

In its Kimi K3 announcement, Moonshot AI says the model has 2.8 trillion total parameters and supports a 1 million-token context window. The company positions Kimi K3 for long-horizon coding, knowledge work, and reasoning tasks, and lists API pricing at $0.30 per million cached input tokens, $3.00 per million cache-miss input tokens, and $15.00 per million output tokens.

Moonshot AI’s Hugging Face model card provides additional implementation details. It describes Kimi K3 as a native multimodal model with 2.8T total parameters, 104B active parameters, and a 1,048,576-token context length. The same model card says the full model weights are available under the Kimi K3 License.

The release places Kimi K3 among a growing set of large open-weight models competing with closed frontier systems, though the precise terms of openness depend on the model license. Tom’s Hardware also reported that Moonshot AI released the Kimi K3 weights and highlighted the company’s claims about performance and cost compared with systems from OpenAI and Anthropic.

What the technical report says about reinforcement learning

The arXiv technical report, titled “Kimi K3: Open Frontier Intelligence,” describes reinforcement learning work across coding-agent domains, including a “coding experience” setting. According to the report, scaling reinforcement-learning FLOPs improves capabilities, and Figure 8 plots both task scores and average assistant steps during RL.

That figure is notable because it separates two quantities that are often discussed together: outcome quality and the amount of interaction needed to reach an answer. In the coding-experience curve described by the report, scores rise as RL compute increases, while the average number of assistant steps eventually declines. A cautious interpretation is that the model may be learning more efficient behaviors in that setting, but the paper’s chart alone does not prove a single cause.

The same report indicates that RL was applied across multiple domains. Differences between domains could reflect the tasks, the reward design, the curriculum, or the available opportunities for shorter successful trajectories. The public sources do not establish that coding is uniquely more optimizable than other domains, so that conclusion should be treated as a question for further study rather than a confirmed finding.

Why the release matters

Kimi K3 is significant for three reasons documented by the available sources. First, Moonshot AI is publishing weights for a very large model rather than offering only a hosted service. Second, the Hugging Face card and company blog state an unusually long context length of 1,048,576 tokens, which is relevant for large codebases, documents, and multi-step workflows. Third, the technical report gives a detailed view into training and reinforcement-learning behavior, including how scores and assistant-step counts change during RL scaling.

For developers and researchers, the most concrete takeaways are the model’s stated scale, open-weight availability, long context window, and the report’s evidence that additional RL compute improved measured capabilities in the evaluated settings. Claims beyond those points, including explanations for why assistant steps fall in some curves but not others, require more controlled evidence than the released chart provides.