AI Briefing
KOSign in

Precision Numerics releases Boole-32B-A2.5B-Preview, a 1.8-bit MoE model 9x smaller than BF16

·2026.11.13 18:29

Key point

Precision Numerics releases Boole-32B-A2.5B-Preview, a 1.8-bit MoE model that is 9x smaller than BF16 and 2x faster than INT4.

Details

  • Precision Numerics released Boole-32B-A2.5B-Preview, an early 1.8-bit version of Recursive's flint model.
  • The model utilizes mixed-precision quantization, expert pruning, and custom kernels.
  • These optimizations make the model 9x smaller than BF16 and 2x faster than INT4.
  • Weights are available on Hugging Face as GGUF, with a demo and a custom llama.cpp fork provided on GitHub.
llmquantizationmoelocal-llmprecision-numericsboole
Read originalShare on Threads

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.

Related

Hacker NewsUnsloth Releases Dynamic V3.1 Quantizations for Qwen3.8-27B and New 1-bit Data Types for Qwen3.8-2.4TRedditUnsloth releases Qwen3.8-27B Dynamic V3.1 GGUFs with 5-10% accuracy gainsRedditUnsloth Releases Dynamic V3.1 GGUFs for Qwen3.8-Flash-Next with 10-15% Accuracy GainsPyTorchKRMistral Large 4 Achieves Coding Performance Similar to GLM-5.3 with 1T MoE ModelPyTorchKRGoogle Sleep, Qwen3-8B Achieves 79.2% on AIME-24Hacker NewsNew Arm NEON kernel speeds up llama.cpp prompt processing by up to 6.5x