Tier 2 Restricted open-weight
What the publisher claims: open weights, efficient, on-device
The one-liner: Efficiency research released under Apache 2.0 but gated on unreleased kernels and an opaque base.
The six questions
weights downloadable, ungated
Apache 2.0
native 1-bit and ternary formats need Prism forks of llama.cpp and MLX; upstream does not support them yet
whitepapers and benchmark methodology published; training data and distillation procedure not open
Limited to moderate.
Prism ML publishes Apache-2.0 weights on Hugging Face, but the 1-bit kernels live in company-controlled forks of llama.cpp and MLX, so the only participation path is filing issues or upstream pull requests that Prism itself writes.
License
Apache 2.0 (HF license:apache-2.0 on Bonsai-8B-gguf, Bonsai-8B-mlx-1bit, Bonsai-8B-unpacked and Ternary-Bonsai releases) OSI-approved — license source ↗
What is actually released
- 1-bit GGUF for llama.cpp
- 1-bit MLX build for Apple Silicon
- full-precision unpacked safetensors
- whitepapers
- benchmark scripts; derived from Qwen3 base weights; 1-bit kernels live in Prism forks of llama.cpp and MLX
Why it matters
Shows that openness of the artefact and openness of the training story can fail independently in the same release.
Sources
- https://huggingface.co/prism-ml/Bonsai-8B-gguf
- https://huggingface.co/prism-ml/Bonsai-8B-unpacked/raw/main/README.md
- https://prismml.com/
Derived from the openness dataset behind But Is It Open? — snapshot 2026-11-07 ↗. Scores are qualitative readings of the cited evidence, not verdicts; tier measures openness, not intelligence.