Moonshot AI

Moonshot AI, created by Yang Zhilin, released Kimi K3 on 16 July 2026: an open-weights model its vendor describes as ~2.8 trillion parameters, a one-million-token context, and a hybrid-attention MoE architecture, with weights to be opened before 27 July. No complete official benchmark table accompanied the launch, so the headline figures (Kimi Delta Attention, decoding announced as 6.3x faster on 1M tokens, +25% training efficiency) stayed vendor-stated and the scores came from community arenas. SFEIR, a partner of Anthropic and Google Cloud, read it as a portfolio question: the pricing reported in first reviews (around $3 per million input tokens, $15 output, $0.30 cached) pulls the price-performance curve down, and open weights add a reversibility column to multi-model routing, at the cost of running 2.8 trillion parameters yourself.

Mozilla's The state of open source AI put Kimi K3 at 57 on the Artificial Analysis Intelligence Index, fourth overall, against 61 for the best closed model, Claude Opus 5.

Two months later the company appears in a different register. Anthropic's threat intelligence report of 10 September 2026 names Moonshot AI among seven Chinese labs doing distillation, describing a cluster (GTG-16002, 5,380 accounts, 23 million exchanges) that served Claude responses to its own customers without telling them and converted the thinking signature into a full trace through cross-session replay. Moonshot AI denies the covert extraction allegation. The organization that raised the open-weights ceiling in July is the one whose customer traffic Anthropic says was silently relayed to Claude.