VQV Signal
According to a post on X by @QwenAIFans, a pending Strata pull request with a new lookup mode got two 2016-era Nvidia P40 GPUs to 55 tokens per second while decoding a 125B Qwen mixture-of-experts model. The post compares this with about 20 tokens per second on Qwen3.8-27B using stock llama.cpp. The change is still in review.
According to a post on X by @QwenAIFans, a pending Strata pull request with a new lookup mode got two 2016-era Nvidia P40 GPUs to 55 tokens per second while decoding a 125B Qwen mixture-of-experts model. The post compares this with about 20 tokens per second on Qwen3.8-27B using stock llama.cpp. The change is still in review.
AI-assisted summary based on the listed source.
A pending Strata pull request with a new lookup mode got two 2016-era P40 GPUs to 55 tokens a second decoding a 125B Qwen mixture-of-experts model, up from about 20 on Qwen3.8-27B with stock llama.cpp. The change is still in review. Why it matters: If it holds up, cheap used server cards become a real option for...
If the result holds up under review, used server GPUs could become a more practical option for people running large open models at home. Developers and hobbyists using llama.cpp-based setups are the most directly affected. Watch for the pull request's review outcome and for independent benchmarks.
Hardware and robotics watchers may want to track whether this becomes a product, benchmark, or deployment signal.
VQV organizes public signals from inspectable sources. It does not independently verify the underlying report.
Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.
VQV surfaced this signal because it is recent, relevant to Open Source LLMs, connected to Scobleizer News Pulse (X).
No login, cookies, social SDKs, or automatic posting.
