2026-10-05 / Signal #7
A 125B model just moved into the gaming PC
“The model that was supposed to live in a warehouse now lives next to a GPU that plays *Cyberpunk*.”
The Story
Strata is an open-source inference stack that runs Qwen 3.8 Flash Next (125B MoE, few experts active per token) on consumer boxes. HN’s headline number is ~100 tok/s on a 4090; the repo’s own tables show high double-digits on a 12 GB card plus 32–64 GB RAM, fully local, OpenAI-compatible localhost. This is a builder-behavior change, not a lab leaderboard.
Why It Matters
Frontier weight classes escaping the datacenter. The cloud’s smokestack vs a tower PC under a desk. Builder utility first, future-shock second.
Evidence
GitHub Niko1221/Strata; HN ~834 pts / ~24h; Linux Compatible.
Sources
Daily scan: 2026-10-05