Third party 'cracks' 5.95GB ternary-compressed Bonsai 2 at the weight level, refusal rate 93.4% to 0%
solyarisoftware · x · 2026-09-23
After PrismML released Bonsai 2, a 5.9GB ternary-compressed Qwen3.8-27B, third-party team dealignai claims it removed refusal behavior at the weight level without growing the file: still 27B params / 5.95GB / 1.7-bit ternary weights, with vision, off/low/xhigh reasoning modes, and CUDA/Metal/CPU support intact. Their own HarmBench-320 test shows refusal dropping from 93.4% to 0%, with MMLU down just 1.23 points (39.69% → 38.46%). Caveat: these are the creators' own tests, not independent benchmarks. The takeaway: models keep shrinking while local operators gain more control.
More from Models
- GPT-6 Sol accused of ignoring configured skills that work fine on GPT-5.6 — Aber-so-richtig · 2026-09-23
- Opus 5.5 ultra wows with 4-agent, 90-minute fully-from-scratch build — repligate · 2026-09-23
- No Kimi release this week, says leaker: Moonshot staff go on holiday Friday — ChrisGPT · 2026-09-23
- Screenshot suggests GPT-5.6 Sol is free for ChatGPT Go subscribers on Windows — sam619007 · 2026-09-23
- Opus 5.5 wins the day as GPT-6 Sol skips ChatGPT, sparking model debate — koltregaskes · 2026-09-23
- Opus 5.5 becomes a daily driver: faster, cheaper than Opus 5, plus reset credits — addyosmani · 2026-09-23