Precision Numerics has debuted with Boole-1.7B, a 1-bit version of Qwen3-1.7B that brings packed weights down to 247 MB, 13.9 times smaller than the BF16 counterpart. The technical figure is not just a curiosity: according to the company's ten-task evaluation, the model retains 85.9% of the BF16 average score, outperforms the 1-bit Bonsai-1.7B, and nearly matches the 1.58-bit Bonsai-1.7B. The weights are available on Hugging Face, the demo code for vLLM and llama.cpp is on GitHub, and a local browser chat is already available.

The move is notable because the comparison is not with a generic compressed model, but with Bonsai, a name already associated with aggressive quantization experiments in the open-source landscape. Precision Numerics does not stop at publishing weights: it points to two usage paths, a local browser demo for low-latency tasks in English and non-thinking mode, and serving with vLLM or llama.cpp. The choice of two widely used inference stacks is not neutral: it reduces friction for those who want to test the model on their own hardware.

The implicit thesis is that 1-bit quantization is becoming a deployment lever, not a laboratory compromise. If a 1.7 billion parameter model compressed to 247 MB retains a significant share of BF16 performance on a task basket, the TCO reasoning changes. Experimenting with a local LLM no longer requires a GPU with large VRAM: the bottleneck shifts to software, latency, and the model's operational constraints. The browser demo, limited to English and non-thinking mode, is a reminder: compression does not remove linguistic and reasoning limits, it makes them more visible.

Who benefits from this announcement? The open-source community and teams managing on-premise or edge services: smaller models reduce memory requirements and simplify pipeline updates. Who might lose? Providers selling cloud-only inference for light workloads: if a 247 MB model runs in a browser, part of the demand for simple-task APIs may migrate to local execution. It is too early to speak of replacement, but the structural signal is there: the race toward low-precision models is accelerating, and the explicit comparison with Bonsai shows that competition is not only about parameters, but about real efficiency after quantization. For those evaluating on-premise deployment, the trade-offs among memory footprint, average quality, and serving stack are the kind of variables AI-RADAR analyzes at /llm-onpremise.

Precision Numerics promises more capable models and asks the community which capabilities it wants to see next. The question is not rhetorical: in a segment where one percentage point of quality and a few megabytes can change deployment, feedback from local users becomes part of the roadmap. Boole-1.7B is a first test bed, but the direction is clear.