Post
65
We're releasing precision-numerics/Boole-1.7B: a 1-bit version of Qwen3-1.7B with 247 MB packed weights.
• 13.9× smaller weight storage than BF16.
• 85.9% of BF16’s average score across our 10-task evaluation.
• +5.58 average points over binary Bonsai-1.7B in the same evaluation.
It was meant for low-latency tasks using non-thinking and English.
Try it locally in your browser, or download the weights and run on your machine.
• Blog: https://precisionnumerics.ai/blog/boole-1p7b
• Code (for vLLM, llama.cpp): https://github.com/PrecisionNumerics/boole-demo
• Local chat in browser: https://precisionnumerics.ai/chat/
We'll continue shipping better models - let us know what models or capabilities you want to see next!
• 13.9× smaller weight storage than BF16.
• 85.9% of BF16’s average score across our 10-task evaluation.
• +5.58 average points over binary Bonsai-1.7B in the same evaluation.
It was meant for low-latency tasks using non-thinking and English.
Try it locally in your browser, or download the weights and run on your machine.
• Blog: https://precisionnumerics.ai/blog/boole-1p7b
• Code (for vLLM, llama.cpp): https://github.com/PrecisionNumerics/boole-demo
• Local chat in browser: https://precisionnumerics.ai/chat/
We'll continue shipping better models - let us know what models or capabilities you want to see next!
