Moonshot AI launches Kimi K3, a 2.8T open-weight model

On July 16, 2026, Chinese AI lab Moonshot AI announced Kimi K3, calling it its most capable model to date with 2.8 trillion parameters. It uses a Mixture-of-Experts (MoE) architecture plus two in-house innovations: Kimi Delta Attention (a hybrid linear attention mechanism) and Attention Residuals. Kimi K3 offers a 1-million-token context window, image input, and an always-on "thinking" reasoning mode.

K3 is available now via kimi.com and the API (model id kimi-k3); an open-weight release is promised "by July 27, 2026" on Hugging Face — expected around 594 GB under the Modified MIT license used for recent Kimi models.

Key points

• A 2.8-trillion-parameter MoE model — roughly 75% larger than DeepSeek V4 Pro — that Moonshot calls the first open "3T-class" model.
• 1M-token context, native vision, and currently only one reasoning effort ("max").
• Priced at $3 per million input tokens and $15 per million output tokens — Claude Sonnet tier, and the most expensive model from a Chinese lab (K2.6 was $0.95/$4).
• On self-reported benchmarks it beats Claude Opus 4.8 and GPT-5.5 but trails Claude Fable 5 and GPT-5.6 Sol; Artificial Analysis ranks it #3–4.
• It leads Arena.ai's Frontend Code arena, ahead of Claude Fable 5; a long-horizon task Elo of 1547 (+732 vs K2.6) while using 21% fewer output tokens.

FAQ

Can I download Kimi K3 weights yet? Not yet — Moonshot promises open weights on Hugging Face by July 27, 2026; for now it runs via the app and API only.

How open is Kimi K3? It is an open-weights model, expected under a Modified MIT license — not a full open-source training release.

Why does it matter? It is the largest open-weight model shipped to date and reaches frontier level on some benchmarks despite China's restricted access to high-end chips.