2.8 trillion parameters, free to download. Kimi K3 is the giant model from Moonshot AI, the Beijing lab backed by Alibaba: it reads images, reasons at every step and loads up to one million tokens of context. Developers and technical teams working on large codebases are its main audience. Access is free on kimi.com, with paid plans from $19 to $199 per month and a token-billed API.
Kimi K3 runs on a Mixture-of-Experts architecture with 2.8 trillion parameters, the largest ever published as open weights. Two in-house inventions, Kimi Delta Attention and Attention Residuals, keep compute costs under control at that scale.
In practice, the model shines on long-haul tasks: it navigates an entire repository, drives terminal tools, fixes bugs and iterates from logs or screenshots. Reasoning stays on at all times, with an effort setting from low to max on the API side.
Moonshot positions its model against the top proprietary systems from Anthropic and OpenAI, and its own benchmarks place it in the top three on most tests. One caveat from independent testing: the model is slower and wordier than the median, at around 62 output tokens per second.
Three reference points on the bill:
The API charges $3 per million input tokens ($0.30 on a cache hit) and $15 per million output tokens, flat across the whole context window. Those figures will shift over time; Moonshot's pricing page is the one to trust.
| Access | Price | Best for |
|---|---|---|
| Kimi app and site | Free | Testing and daily use |
| Memberships | $19 to $199 per month | Heavy agentic sessions |
| API | $3 in, $15 out per M tokens | Developers and products |
| Self-hosting | Free weights, hardware on you | Companies with GPU infrastructure |
Yes, Kimi K3 is free through the Kimi website and app. Paid tiers add agent credits, higher concurrency and extra-long chats up to one million tokens of context. The API is billed separately, per million tokens consumed.
Technically yes, since the weights are public under a Modified MIT license, but not on a personal machine. Moonshot recommends at least 64 GPU accelerators to serve the model, so self-hosting is a job for companies with real infrastructure, which then keep full control of their data.
Your conversations pass through Moonshot's servers, on Chinese infrastructure, and the privacy policy states that content may be used to train the models, with no documented opt-out in the consumer product. For sensitive material, self-hosted weights remain the safer route.
Verdict: Kimi K3 fits developers and teams who want frontier-level help on large codebases without depending on the American labs. For sensitive data, self-hosting keeps everything under your control.
