Independent scan: 185 of 265 open LLMs can infer their training GPU family
willdepue · x · 2026-11-14
An independent reproduction confirms willdepue's finding: scanning 265 open LLMs, 185 can infer the GPU family they were trained on, leaking pretraining hardware details.
- Confirms all original findings
- Corrects one detail: Qwen3-235B's April checkpoint reported CPU initialization because unused rows were overwritten with CPU noise after training, while Qwen's 2507 releases keep rows initialized on Hopper
- Some models were even pretrained on AMD hardware
More from Models
- Claude iOS hit by Sign in with Apple outage affecting new signups — ClaudeAI-mod-bot · 2026-11-14
- How credible are frontier labs' claims of solving Navier–Stokes and Fermat's Last Theorem? — Witty-Ad-8579 · 2026-11-14
- Users pile on OpenAI: buggy desktop app and DevDay flop push them to Claude — koltregaskes · 2026-11-14
- Heavy AI users stuck between Pro and Ultra tiers as daily usage lands mid-tier — gaganghotra_ · 2026-11-14
- Google Search reportedly rolling out Gemini 4 Hydrogen globally with AI-first results — feblet · 2026-11-14
- Dev spent up to $2,000/month on AI Pokémon fan game — Claude now refuses to help — tekbog · 2026-11-14