Moonshot AI opens Kimi K3's full weights on July 27 — the largest open model to date

Moonshot AI is set to release Kimi K3's full weights under a Modified MIT license on July 27, 2026 — making it the largest open-weight model shipped to date, at 2.8 trillion parameters, MarkTechPost reports. The model itself has already been live via Kimi's API and apps since July 16.

What the model actually is, technically

• 2.8 trillion parameters in a mixture-of-experts (MoE) architecture — but only 16 of its 896 experts activate per token, meaning roughly 50 billion parameters actually run per forward pass.
• Custom architecture — Kimi Delta Attention (KDA) and Attention Residuals (AttnRes), which the developers say deliver roughly 2.5x efficiency gains over the prior K2 model.
• A 1-million-token context window with native text, image, and video support.
• "Thinking mode" is on by default — the model reasons through problems without needing a separate reasoning variant.

How it performs on benchmarks

According to Interconnects, Kimi K3 leads on BrowseComp, SWE Marathon, and OmniDocBench, but trails Claude Fable 5 on FrontierSWE and HLE-Full. Its Intelligence Index score of 57.11 puts it fourth among frontier models — behind Fable 5 and GPT-5.6 Sol, but ahead of Claude Opus 4.8 — at roughly a quarter of their API cost.

The technical story, apart from the politics

Whatever happens with the dispute over the model's origins — we covered the White House's accusation that Moonshot AI illegally distilled Anthropic's model earlier — Kimi K3 has already reached the market as a real, functioning product with open weights. Releasing the full weight set on July 27 will let independent researchers examine the model's architecture for themselves.

Nothing here should be taken as financial advice — just information to consider.