Moonshot AI Ships Kimi K3: 2.8 Trillion Parameters and a 1M-Token Context Window
Chinese lab Moonshot AI launched Kimi K3 on July 16, 2026—a 2.8T-parameter MoE flagship with a 1-million-token context window, native vision, and open weights promised by July 27—intensifying the global frontier-model race.
A quiet launch of a loud model
Moonshot AI began rolling out Kimi K3 on July 16, 2026, with little of the traditional launch-day theater. The model appeared first on Kimi Code, the Kimi app, and Moonshot’s OpenAI-compatible API—then quickly became a reference point in conversations about open-weight scale.
What K3 brings
According to Moonshot’s documentation and early technical coverage:
- ~2.8 trillion total parameters (Mixture-of-Experts)
- 1,048,576-token (1M) context window
- Native multimodal input (text, image, video) with text output
- New architecture pieces including Kimi Delta Attention and Attention Residuals
- Always-on reasoning mode aimed at long-horizon coding and agent work
Moonshot has described K3 as the world’s first open model in the 3-trillion-parameter class, with full weights promised by July 27, 2026. Until those weights land, K3 is primarily an API-era flagship—already live on OpenRouter and Kimi’s own platforms.
Competitive context
The release lands in the same news cycle as OpenAI’s Hugging Face security disclosure and ongoing U.S.–China competition over frontier systems. For builders, the near-term question is practical: can a 2.8T open model be served, quantized, and adapted outside Moonshot’s stack? For geopolitics and open-source strategy, July 27 is the date that matters.
Key takeaways
- K3 is live as an API product today; full open weights remain a dated commitment (July 27, 2026).
- Public specs emphasize MoE scale (~2.8T), a 1M-token window, and native multimodal input.
- Coverage treats the launch as both a capability and distribution story—API reach now, weight availability next.