Unknown · 3× · tolest 20. Sept. 2026
38
Momentum
Bonsai 1.7B is a native 1-bit language model by PrismML (a startup emerging from Caltech research) based on the Qwen3 architecture. The weights are quantized to a single sign bit per parameter (effectively 1.125 bits/weight including scaling factors), resulting in a GGUF file size of only around 248-290 MB. It is part of the broader Bonsai model family (1.7B, 4B, 8B, 27B) and runs locally on laptops, smartphones, and even in the browser via WebGPU thanks to its extremely low memory footprint. The model is provided free of charge under the Apache-2.0 license.
Momentum-Verloop
22.06.20.09.
Features
| Key Benchmark (%) | GSM8K evaluation linked via official Hugging Face leaderboard (concrete score viewable there, not stated in text form on model card) |
| Context Window (Tokens) | Based on Qwen3 architecture (context length for 1.7B variant not separately specified officially; larger Bonsai models use 262,144 tokens) |
| License | Apache 2.0 (free use including commercial) |
| Multimodality | Text-only (no image input); vision capabilities only available starting from the 27B models in the Bonsai family |
| Platform | Local via llama.cpp (GGUF, Q1_0) on CPU/GPU (CUDA, Metal), MLX for Apple Silicon/iPhone, and WebGPU in-browser |
| Release Date | March 31, 2026 |