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
- Theory: Typos Disrupt LLM Tokenization, Which May Explain Why AI Text Has a Distinct Smell — menhguin · 2026-11-11
- LMArena co-founder explains AI evaluation methodology and policy in new podcast — MattPerault · 2026-11-11
- Runway open-sources Praxis-1, turning video pretraining into real robot control — runwayml · 2026-11-11
- DeepMind launches AlphaGenome 2, spotting nearly 2x more gene-regulating mutations — GoogleDeepMind · 2026-11-11
- MazeBench: Fable 5.2 scores 15%, near-frontier at half price; Gemini 4 Carbon 11% — patience_cave · 2026-11-11
- ERNIE 5.5 Flash Preview enters LMArena Battle Mode for crowdsourced ranking — arena · 2026-11-11