American open-source labs race to counter China's AI models

As Chinese open-weight models (Qwen 3.8, Kimi K3, GLM-5.2, DeepSeek V4) surge, Forbes reports a countermovement among Western open-source labs. Mistral made over $200M last year selling open-weight models to Western enterprises and governments, leaning on a 'not Chinese and not American' neutrality pitch that appeals to buyers wary of both U.S. cloud lock-in and Chinese provenance. U.S. startups including Poolside — which released its open model in the Laguna line — are racing to build credible homegrown open alternatives, potentially aided by a Trump administration eager to blunt Chinese AI influence.

The strategic logic is that open weights have become a geopolitical instrument: whoever's open models developers build on gains ecosystem gravity. Chinese labs have moved aggressively into that space, and the U.S. response is fragmented between labs lobbying to restrict open source (see the reported ban debate) and startups arguing the answer is more competitive American open models, not fewer. NVIDIA underscored the latter camp, publicly congratulating Poolside on Laguna S 2.1 — 'open-weight, delivers way beyond its size, runs great locally' and customizable with NVIDIA NeMo — and Perplexity's Aravind Srinivas tweeted simply 'American open source frontier.'

Competitively this is the week's meta-theme wrapping the HF incident, the Chinese model wave, and the ban debate. Skeptics note U.S. open models still trail the best Chinese ones on price-performance, and that policy uncertainty (a possible open-source ban) undercuts the very startups trying to compete. Watch Poolside's benchmark reception and whether Washington's policy lands on restriction or promotion.