Moonshot's Kimi K3 lands as largest open-weight model yet, triggering a second DeepSeek shock

The Beijing lab's 2.8-trillion-parameter MoE — with a 1-million-token context window and full weights due July 27 — beat Claude Opus 4.8 and GPT-5.5 on coding and general-agent tasks, and sent Chinese AI rivals and US chip stocks sliding on Friday.

Beijing-based Moonshot AI released Kimi K3 late Thursday, a 2.8-trillion-parameter sparse mixture-of-experts model with a 1-million-token context window, and by Friday’s Asian open the reaction was already legible in the tape: Z.ai down nearly 30% in Hong Kong, MiniMax off 16%, TSMC down 7%, SoftBank down 9%, Nvidia off 1.2%. The full weights are scheduled to ship July 27, which means the largest open-weight model ever released is also a public artifact anyone can host.

On Arena.AI’s independent leaderboard, K3 edged past Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5 on coding and general-agent tasks, including Program Bench and SWE Marathon. It doesn’t beat everything. Moonshot itself concedes the model “still trails the most powerful proprietary models,” a reference to Claude Fable 5 and GPT-5.6 Sol. But the price curve is the story the market is pricing. K3 runs at $3 per million input tokens and $15 per million output, with cached input at $0.30. Fable 5 charges $50 per million output.

The architectural claims are specific enough to matter: a hybrid linear-attention mechanism Moonshot calls Kimi Delta Attention, plus a technique labeled Attention Residuals. Bank of America’s Alex Liu read the release as evidence that “pre-training scaling, paired with architectural innovation, can still deliver step-change gains,” and that “K3 raises the capability ceiling for China AI models, shifting the burden of proof to other independent AI labs.”

Patrick Moorhead of Moor Insights called Friday’s selloff “an over-reaction shockingly similar to the DeepSeek panic,” and he’s structurally right. TSMC just posted a 77% jump in quarterly operating profit; Apple’s supply chain wasn’t rerated on fundamentals. What was rerated is the narrative that only US labs, backed by hyperscaler capex, can hold the frontier.

Moonshot, backed by Alibaba and Tencent, raised $2 billion in May at a $20 billion valuation and is now reportedly raising at roughly $30 billion. Xi Jinping’s industrial-policy read on open-source AI as strategic infrastructure has, for the second time in eighteen months, been vindicated by a single Friday.

Moonshot's Kimi K3 lands as largest open-weight model yet, triggering a second DeepSeek shock

Sources