Moonshot AI releases Kimi K3, a 2.8T-parameter open-weights multimodal model
Confirmed
Importance: historic (5 of 5)Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- Historic (5 of 5)
- Last verified
- 11 October 2026
Your AI and this story
- GPT-6 Astra77 days after its cutoff
- Claude Opus 5.516 days after its cutoff
- Gemini 4 Hydrogen107 days after its cutoff
- Grok 4.746 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 16 days before it.
Key facts
- 2.8T total parameters, ~104B activated (16 of 896 experts per token + 2 shared) per Hugging Face model card
- Context window: 1,048,576 tokens; 401M-parameter MoonViT-V2 vision encoder; weights released in MXFP4 with MXFP8 activations
- Architecture: Kimi Delta Attention + Gated MLA layers, Stable LatentMoE, Attention Residuals
- Model card benchmarks: GPQA Diamond 93.5, BrowseComp 91.2, Terminal-Bench 2.1 88.3, DeepSWE 67.5, Video-MME 90.0
- API pricing: $3/M input, $15/M output (vs $50/M output for Claude Fable 5 cited by Fortune)
- ARC Prize: 94.5% ARC-AGI-1, 60.4% ARC-AGI-2
- License: custom Kimi K3 License (separate agreement for MaaS businesses >$20M revenue; attribution above 100M MAU)
- Listed on Amazon Bedrock 2026-09-18 (secondary report)
Show 2 more
- US reaction (Fox Business, July 21, 2026): Treasury Secretary Scott Bessent, with Kimi K3 in the news: ‘This administration supports open source models, but what we do not support is IP theft. If we see, especially, that overseas models are stealing from our great companies, we have the ability to sanction them’ (Gizmodo)
- July 22, 2026: OSTP director Michael Kratsios said on X that Moonshot ‘distilled Anthropic’s Fable for the development of its K3 model’ and accessed GB300 servers in Thailand, without public evidence (entry White House science chief Kratsios accuses Moonshot of distilling Claude Fable to build Kimi K3 and of using GB300 servers in Thailand)
What happened
Moonshot AI launched Kimi K3 on 2026-07-16 as a native multimodal, agentic flagship. The Hugging Face model card lists 2.8T parameters with ~104B active, a 1M-token context, and MXFP4 weights produced with quantization-aware training. Fortune (which gave 2.7T) reported Moonshot’s claims of being competitive with Anthropic’s Claude Fable 5 and substantially outperforming Claude Opus 4.8 and GPT-5.5, particularly at long-running coding sessions and terminal tool orchestration. Weights followed on Hugging Face by late July. Moonshot also claimed an official 42/42 on IMO 2026 problems (per commentary quoted by TechXplore; not independently confirmed here).
Why it matters
K3 made the “open-weights frontier” roughly one step behind the very best closed models, at a fraction of their price, and the weights are downloadable by anyone — a major data point in the US-China model race and for open-model policy debates.
Sources
5 sources from 5 sites. Numbers match the chips in the text.
5 sources: 1 primary, 3 press, 1 reaction
Primary
- Hugging Face: moonshotai/Kimi-K3 model cardhuggingface.co, official
Press
- Gizmodo (July 21, 2026): US Treasury chief threatens sanctions on Chinese AI labs over IP theft concernsgizmodo.com, press
- Fortune: Kimi K3 pushes Chinese AI into Fable-level territoryfortune.com, press
- Bloomberg: Moonshot unveils Kimi K3, narrowing gap with US rivalsbloomberg.com, press
Reactions
- ARC Prize resultsarcprize.org, discussion
Changes
- Linked the UK AISI/CAISI cyber assessment (July 23) and the Frontier Security eval-cheating report (Aug 8), both own entries
- Added the July 22 Kratsios distillation accusation (own entry)
- Added Bessent’s July 21 sanctions warning (found via a re-dated Stocktwits story)
- Filed
