Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-07-21 · high
Moonshot's 2.8 trillion parameter model tops Claude on a blind coding eval the same day Bessent floats sanctions on Chinese AI.
Why now: Full weights release July 27, and the sanctions talk broke the same day as the benchmark news, per Bessent's July 21 comments to Fox Business.
Moonshot AI released Kimi K3, a sparse mixture-of-experts model activating just 16 of 896 experts per token, about 1.8% of the pool live at any moment. It has a 1M-token context window, native vision, and two new architectural pieces called Kimi Delta Attention and Attention Residuals. Full weights land July 27 under a modified-MIT-style license, with hosted pricing at $0.30/M cache-hit input, $3/M cache-miss, $15/M output.
Here's the part that should actually change what you build with. Kimi K3 ranked first on Arena's Frontend Code evaluation at 1,679 points, ahead of Claude Fable 5 in blind developer testing. On overall capability it still trails Fable 5 and GPT-5.6 Sol. So this isn't China catching up across the board, it's narrower than that: on the specific job of writing frontend code, an open model with published weights beat the model I use every day in my personal projects.
That news landed the same day Treasury Secretary Scott Bessent told Fox Business the administration will investigate whether Chinese models were distilled from US systems, saying "we are finding watermarks of our U.S. large language models on many of the Chinese models, and that's unacceptable," with sanctions on the table. This follows Anthropic's June 10 letter to Senate Banking alleging Alibaba's Qwen lab ran 28.8 million Claude exchanges through roughly 25,000 fraudulent accounts between April 22 and June 5.
MIT Technology Review reports the White House AI apparatus is split: one camp wants export controls on cheap capable Chinese weights, another thinks the bigger risk is the US retreating from open models. CAISI lost its director three months in (Chris Fall resigned July 20, NIST's Arvind Raman is acting), and a voluntary 30-day classified pre-release review framework with OpenAI, Anthropic, and Google is still being finalized.
I'm not a lawyer. But weights on your own disk survive a policy fight that a hosted API subscription doesn't.
Each link below shares sources, entities, or timing with this story.
Anthropic released Fable / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (Anthropic released Fable); both cover Alibaba, Anthropic, Chinese, Fable; overlapping topics (chinese, july, model, weight).
Anthropic released Fable / Shared entities / Same source domain / Earlier coverage
Linked by a graph relationship (Anthropic released Fable); both cover Alibaba, Anthropic, April, Full; reported by the same outlet (venturebeat.com).
Anthropic released Fable / Shared entities / Same source domain / Shared topic / Earlier coverage
Linked by a graph relationship (Anthropic released Fable); both cover Alibaba, Anthropic, April, Chinese; reported by the same outlet (cnbc.com).
Linked by a graph relationship (Anthropic released Fable); both cover Anthropic, Chinese, CLAUDE, GPT; reported by the same outlet (cnbc.com).
Anthropic released Fable / Shared entities / Earlier coverage / Tension
Linked by a graph relationship (Anthropic released Fable); both cover Alibaba, Anthropic, Chinese, Google; earlier Alibaba coverage from 2026-04-20.
Anthropic released Fable / Shared entities / Same source domain / Shared topic / Earlier coverage / Tension
Linked by a graph relationship (Anthropic released Fable); both cover Anthropic, Claude, Fable, OpenAI; reported by the same outlet (axios.com).
Anthropic released Fable / Shared entities / Same source domain / Shared topic / Earlier coverage
Linked by a graph relationship (Anthropic released Fable); both cover Anthropic, Chinese, Claude, Google; reported by the same outlet (axios.com, cnbc.com).
Anthropic released Fable / Shared entities / Shared topic / Earlier coverage
Linked by a graph relationship (Anthropic released Fable); both cover Anthropic, Claude, Google, GPT; overlapping topics (anthropic, claude, code, model).