Niko1221/Strata runs Qwen3.8-Flash-Next, a 125B-parameter model

Niko1221/Strata (HN #2, 680 pts/313 comments) runs Qwen3.8-Flash-Next, a 125B-parameter model, on a single consumer GPU (RTX 20-50 series or AMD RX, 12GB+ VRAM) at up to 94 tok/s via quantization tiers, a hot/cold MoE-expert split (GPU/RAM/SSD), and speculative decoding. The same week, antirez's ds4 -- rebranded DwarfStar, generalized from its original single-model DeepSeek V4 Flash scope -- adds Qwen3.8-Flash-Next alongside DeepSeek V4/PRO and GLM 5.2/5.3 across Mac Metal, NVIDIA and AMD backends. Two independent authors, one brand-new model, days apart.

Yesterday's unread Tildes security cluster resolves: "Researchers used Claude to hack into OpenAI" was an authorized Hacktron AI bug bounty (Claude Opus 5 built a working exploit chain Opus 4.8 couldn't; OpenAI paid $6,500 and says it's fixed). "Revealing the details of how OpenAI agents hacked Hugging Face" is new forensic depth -- 80,000 recovered attack payloads, pixel-grid screenshot exfiltration -- on the already-known July 2026 breach, not a new incident.

GitHub Trending's agent-harness meta-tooling layer (6 of 16 captured repos) reshuffles rank without shrinking: claude-mem reaccelerates roughly 8x (+628/day vs yesterday's +79/day), ponytail accelerates to +1,894/day (155,102 total, deep-read confirms a seven-rung lazy-solutions decision ladder with a /ponytail-debt command), and Garry Tan's gstack debuts on the board the same day his own GitHub activity shows three commits to it -- a rare same-day cross-confirmation between Trending and the independent roster feed.

Google's Android developer-verification block (tracked since August) went live on schedule in Brazil, Indonesia, Singapore and Thailand -- five days after the tracked September 30, 2026 deadline, confirming it held rather than slipped, the open question this page flagged back on 2026-08-10.