Unknown · depuis 16. Juli 2026 (API-Launch); 27. Juli 2026 (offene Gewichte) · 3× · vu le 27 sept. 2026
Kimi K3 is Moonshot AI's current flagship language model, an open-weight Mixture-of-Experts model with 2.8 trillion total parameters (104 billion active per token), built on the new Kimi Delta Attention architecture with Attention Residuals. It offers a context window of up to 1,048,576 tokens (~1M), native multimodal input (text, image, video), and targets coding, agentic workflows, and scientific reasoning. The model launched via API on July 16, 2026, with full weights following on July 27, 2026 under a custom "Kimi K3 License" (MIT-like with commercial usage gates). On the GPQA Diamond benchmark, Kimi K3 officially scores 93.5%, the best published score for an open-weight model to date.
Fonctionnalités
| Key Benchmark (%) | 93.5% on GPQA Diamond (best known score for an open-weight model) |
| Context Window (Tokens) | 1,048,576 tokens (1M) |
| License | Custom "Kimi K3 License" (MIT-like, with revenue/user thresholds for commercial MaaS offerings) |
| Multimodality | Native visual understanding (text, image, video) |
| Platform | Kimi.com, Kimi Code, Kimi API (platform.kimi.ai), Hugging Face, OpenRouter, Fireworks, Together AI, Baseten |
| Price per 1M Tokens | $3.00 input (cache-miss) / $0.30 input (cache-hit) / $15.00 output (Moonshot API) |
| Release Date | July 16, 2026 (API launch), weights on July 27, 2026 |