---
license: apache-2.0
pipeline_tag: text-generation
library_name: boole-demo
base_model: Qwen/Qwen3-1.7B
base_model_relation: quantized
language:
- en
tags:
- qwen3
- 1-bit
- quantized
- safetensors
- vllm
- humming-kernels
---

<p align="center">
  <img src="./assets/logo-white-on-black.png" width="128" height="128" alt="Boole-1.7B logo">
</p>

<h1 align="center">Boole-1.7B</h1>

<p align="center">A small and strong 1-bit LLM</p>

<p align="center">
  <a href="https://huggingface.co/precision-numerics/Boole-1.7B">Safetensors</a>
  &nbsp;&middot;&nbsp;
  <a href="https://huggingface.co/precision-numerics/Boole-1.7B-GGUF">GGUF</a>
  &nbsp;&middot;&nbsp;
  <a href="https://precisionnumerics.ai/chat/">💬 On-Device Chat</a>
  &nbsp;&middot;&nbsp;
  <a href="https://precisionnumerics.ai/blog/boole-1p7b">Blog</a>
  &nbsp;&middot;&nbsp;
  <a href="https://github.com/PrecisionNumerics/boole-demo">Code</a>
</p>

Boole-1.7B is a 1-bit language model derived from Qwen3-1.7B. Its weights take 247 MB,
13.9× less than the BF16 original, and it keeps 85.9% of the original's ten-task
benchmark average (50.55 vs 58.83).

![Boole-1.7B against Bonsai-1.7B (1-bit and ternary), 4-bit LFM2.5 (QAD) and Granite-4.0-H-350M, on the ten-task average, each benchmark and stored weights](https://huggingface.co/precision-numerics/Boole-1.7B/resolve/main/assets/boole-under-500mb.svg)

## Files

| File | Runtime | Size |
| --- | --- | --- |
| [Boole-1.7B.safetensors](https://huggingface.co/precision-numerics/Boole-1.7B/blob/main/Boole-1.7B.safetensors) | vLLM + Humming · CUDA | 247.1 MB |

For llama.cpp, use the separate [GGUF release](https://huggingface.co/precision-numerics/Boole-1.7B-GGUF).

## Quick start

This is a packed 1-bit checkpoint. Use the [boole-demo](https://github.com/PrecisionNumerics/boole-demo) loader to export it for vLLM + Humming on a CUDA GPU. The install pins vLLM 0.31.0 and humming-kernels 0.1.16.

```bash
git clone https://github.com/PrecisionNumerics/boole-demo.git && cd boole-demo
python -m pip install -e '.[vllm]'
boole-prep vllm precision-numerics/Boole-1.7B --out out/humming
vllm serve out/humming \
    --dtype bfloat16 --max-model-len 4096 \
    --default-chat-template-kwargs '{"enable_thinking": false}'
```

The loader downloads the checkpoint, config, and tokenizer files. See the [runtime README](https://github.com/PrecisionNumerics/boole-demo#vllm--humming) for a chat request example.

## Intended use

English-language inference in non-thinking mode (`enable_thinking=false`).

## Evaluation

Scores are the equal-weight mean of ten tasks: MMLU-Redux, GPQA-Diamond, MuSR, IFEval,
IFBench, GSM8K, MATH-500, HumanEval+, MBPP+ and BFCLv3, run in non-thinking mode with
greedy decoding (GPQA-Diamond averages ten sampled runs). Harness: EvalScope 1.4.2 and
bfcl-eval 2025.10.27.1. See the [blog](https://precisionnumerics.ai/blog/boole-1p7b) for the comparison.

<details>
<summary>Full evaluation settings</summary>

We use the same evaluation settings as [PrismML's Bonsai](https://prismml.com/news/bonsai-8b). The settings below cover 10 generative benchmarks.

| Setting | Value |
| --- | --- |
| Few-shot examples | 0 |
| Generative prompts | Benchmark prompts with the checkpoint's chat template |
| Thinking mode | Disabled (`enable_thinking=false`) |
| Default generation | Greedy: temperature 0, top-p 1 |
| Default seed / repetitions | Seed 42; one response per example |
| Additional sampling filters | Top-k and min-p disabled; repetition penalty 1.0 |
| Dataset limit | Full benchmark splits; no sample limit |

GPQA-Diamond uses temperature **0.6**, top-p **0.95**, and **10 repetitions per
example** with seeds **42–51**. Its score is averaged across those repetitions.

Maximum generated tokens per response:

| Benchmark | Max new tokens |
| --- | --- |
| MMLU-Redux | 2,048 |
| GPQA-Diamond | 8,192 |
| MuSR | 2,048 |
| IFEval | 4,096 |
| IFBench | 4,096 |
| GSM8K | 2,048 |
| MATH-500 | 8,192 |
| HumanEval+ | 8,192 |
| MBPP+ | 4,096 |
| BFCL-v3 | 4,096 |

Generation context is 16,384 tokens by default and 40,960 for BFCL-v3, capped
when a checkpoint specifies a lower limit. Concurrency and memory allocation
are tuned to the runtime and available GPU memory.

IFEval and IFBench report the mean of strict prompt-level and instruction-level
accuracy. HumanEval+ and MBPP+ use augmented tests and pass@1. BFCL-v3 reports the
macro average of 13 single-turn subsets, with `is_fc_model=false` and
`underscore_to_dot=false`. GSM8K and MATH-500 use a 10-second per-answer scoring
deadline; timed-out answers receive zero.

Evaluation packages: **EvalScope 1.4.2** and **bfcl-eval 2025.10.27.1**.

</details>

## License

Weights and runtime code use [Apache 2.0](https://huggingface.co/precision-numerics/Boole-1.7B/blob/main/LICENSE).
[NOTICE](https://huggingface.co/precision-numerics/Boole-1.7B/blob/main/NOTICE) includes upstream Qwen attribution.

## Citation

```bibtex
@misc{precisionnumerics2026boole1p7b,
  author = {{Precision Numerics}},
  title = {{Boole}-1.7B},
  year = {2026},
  month = {October},
  url = {https://huggingface.co/precision-numerics/Boole-1.7B}
}
```