Open Source · JULY 23, 2026
Moonshot's Kimi K3 lands at 2.8 trillion parameters, weights follow July 27
Moonshot AI unveiled Kimi K3 on July 17 — a 2.8T-parameter sparse MoE with a 1M-token context — and promised full open weights ten days later. The release reshuffled the open-weight frontier and knocked TSMC down 7%.
Moonshot AI unveiled Kimi K3 on Friday, a 2.8-trillion-parameter sparse mixture-of-experts model with a 1M-token context window, and committed to publishing full weights on July 27. That ten-day gap between announcement and download is the whole story: the largest openly redistributable model ever released, staged so that traders had to price it before anyone outside Beijing could run it.
The market did its usual thing. TSMC closed down 7% the same day it reported a 77% jump in quarterly operating profit. SoftBank fell 9%. Nvidia slipped 1.2%, the Nasdaq 100 lost roughly 1%, and Hong Kong-listed Chinese competitors caught the worst of it, with Z.ai down 28% and MiniMax off 16%. Bloomberg's read is that the sell-side is now watching memory suppliers like SK Hynix rather than pure compute, on the theory that K3's 1M-token context makes this more a memory story than a GPU one.
K3 is roughly 75% larger than the prior ceiling for Chinese open releases, which had settled around the 1.6-trillion-parameter mark occupied by Meituan's LongCat-2.0 and DeepSeek V4-Pro. On third-party evaluations tracked by Artificial Analysis, Arena.ai and Vals AI, it ranks first on blind front-end coding and second overall behind Anthropic's Claude Fable 5, ahead of OpenAI's GPT-5.6 Sol. Moonshot's own framing is careful. The model, the company says, "demonstrated frontier-level performance across our evaluation suite" but "still trails the most powerful proprietary models, Claude Fable 5 and GPT 5.6 Sol." That's the correct disclaimer to file with regulators and the incorrect one to file with markets, which read the benchmark tables and not the caveats.
The architecture pitch, per Moonshot's release, is "two significant architectural upgrades that improve computing efficiency and enable it to complete long-horizon coding tasks with minimal human supervision." Translated: they're claiming DeepSeek's efficiency lesson generalizes upward, not just downward.
Alex Liu at Bank of America put the strategic read plainly. "K3 raises the capability ceiling for China AI models, shifting the burden of proof to other independent AI labs." He added that "despite persistent hardware/compute capacity constraints in China, K3 demonstrates that pre-training scaling, paired with architectural innovation, can still deliver step-change gains for flagship Chinese models."
The DeepSeek R1 shock of early 2025 lopped roughly $600 billion off Nvidia's market cap in a single session and taught the market a template. Friday's selloff was that template running a second time, faster and with less conviction. Moonshot, founded in 2023 and backed by Alibaba and Tencent, raised $2 billion in May at a valuation above $20 billion; Bloomberg reports a fresh $2 billion round targeting $30 billion is already in the works.
Ten days until the weights land. The pricing has already happened.
Sources
- https://finance.yahoo.com/technology/ai/articles/chinas-moonshot-unveils-worlds-largest-020622030.html
- https://www.bloomberg.com/news/articles/2026-07-20/moonshot-s-kimi-k3-may-be-more-about-memory-than-compute
- https://www.cnbc.com/2026/07/17/moonshot-ai-kimi-k3-model-openai-anthropic-china.html
- https://fortune.com/2026/07/17/china-moonshot-kimi-k3-markets-china-ai/
- https://techcrunch.com/2026/07/18/kimi-threat-or-menace/