← Back to feed

Almost 200 Silicon Valley companies, including Y Combinator portfolio firms and Proton, have signed a letter to the Trump administration arguing that restricting access to Chinese open-weight AI models would cripple the next generation of US startups. The letter arrived as the White House separately accused Moonshot AI of conducting large-scale covert distillation of Anthropic’s Fable 5 to train the very model the signatories want access to.

The model at the centre of both events is Kimi K3: a 2.8-trillion-parameter open-weight multimodal model released by Moonshot AI on July 17. It entered the Artificial Analysis Intelligence Index at 64 points — within striking distance of Claude Fable 5 — and took the frontend code Arena leaderboard at Elo 1679, ahead of both Fable 5 and GPT-5.6 Sol. It is the largest open-weight model published and costs $3 input/$15 output per million tokens when accessed through Moonshot’s API, or zero if self-hosted.

The Commercial Dependency Argument

The founders’ letter focuses on build costs. A startup that trains or fine-tunes on top of Kimi K3 avoids the $3-10 per million token cost of calling Fable 5 or GPT-5.6. For products that route tens of millions of tokens daily, the difference is existential to margin. Blocking access to Chinese open-weight models would effectively mandate frontier spending at US lab prices, concentrating the economic benefits of AI further toward OpenAI, Anthropic, and Google.

Y Combinator has made variants of this argument before, most recently in its filing against the Biden administration’s AI chip export controls. The new letter is sharper: it names specific models, cites performance parity with US frontier systems, and argues that the national security framing conflates model weights with controlled military hardware.

The White House Counter-Position

The Trump administration’s position, stated publicly by White House OSTP director Michael Kratsios, is that Moonshot conducted systematic covert distillation of Anthropic’s Fable 5 to build Kimi K3. The specific claim: Moonshot developed an internal platform to conduct large-scale API extraction while rotating access methods to avoid detection. The company also acquired GB300-equipped servers in Thailand, likely to run the training.

The legal framing is trade secret theft rather than national security export control — a distinction that matters for what remedies are available. Commerce Department entity listing, Treasury investment restrictions, and civil litigation are all on the table. Criminal referral is possible but harder to execute with a China-based defendant.

Moonshot has not responded publicly to the distillation allegation. Rohan Paul’s analysis on Nitter noted a structural evidentiary problem: calling the Fable API does not prove Fable generated the response, because Anthropic has stated it redirects suspected distillation requests to Opus rather than serving Fable. Proving Kimi K3’s training corpus came from Fable specifically requires served-model IDs, corpus-to-training linkage, and model-lineage evidence such as canary transfer or unique error fingerprints.

The Split Running Through the Debate

OpenAI’s Dean Ball framed the Kimi K3 release as a reason to reconsider open-weight AI policy broadly, calling the model’s capabilities a strategic threat. TechCrunch and multiple independent commentators noted that the framing conflates two separate issues: the economic competitiveness of US AI labs, and whether model weights themselves constitute a weapons-class security risk.

The letter from 200 founders represents the most organised industry pushback to the security framing. The signatories’ argument is not that Kimi K3 poses no risk — it is that restricting US developer access to the weights would do more damage to the US AI ecosystem than it would prevent from adversary development.

The administration has not announced a specific restriction. The letter is pre-emptive, aimed at influencing a decision that has not yet been made but that sources in Washington describe as under active consideration.

Numbers at Stake

Kimi K3 open weights: 2.8T parameters, available for download. Self-hosted inference at 2-bit quantisation runs on a single high-memory workstation. The model scores 93.4% on SWE-bench Verified, 90.7% on LiveBench Reasoning, and leads Harvey LAB-AA legal tasks at 26.7%, nearly double Fable 5. For any developer building a product where legal, coding, or reasoning capability matters, the performance gap with US-only alternatives is no longer significant.

The policy outcome will turn on whether the administration frames the question as an export control matter — where models weights could be treated like controlled technology — or as a trade secret issue requiring enforcement against Moonshot specifically rather than restrictions on US developers.