1-bit Boole-1.7B packs Qwen3-1.7B into 247MB and beats Bonsai-1.7B
d77chong · reddit · 2026-10-12
Precision Numerics launched Boole-1.7B, an open-source 1-bit quantization of Qwen3-1.7B:
- 247MB packed weights, 13.9× smaller than BF16
- Retains 85.9% of the BF16 model's average score across a 10-task eval, beating 1-bit Bonsai-1.7B and nearly matching its 1.58-bit version; GSM8K 74.7, MATH-500 49.8, HumanEval+ 53.7
- Works with vLLM and llama.cpp, plus a browser-based local demo
Weights, demo code and a blog post are all public; the team says more capable open-source models are coming.
More from Models
- SubQ Snap launches: 384K-token decision model with constant per-token compute — ryanallred · 2026-10-13
- Dev hands-on: Muse clearly beats Dots on tool calling, personalization feels natural — evijit · 2026-10-13
- Meta's own Muse assistant insists Muse Wonder doesn't exist hours after launch — msg · 2026-10-13
- Muse Wonder trails Opus on coding, but gives Meta a real base model — zephyr_z9 · 2026-10-13
- Subquadratic launches new model, first 10K signups get 100M free tokens — ashleymayer · 2026-10-13
- Meta launches Muse Wonder, a new frontier model family — thisiskp_ · 2026-10-13