News · 2026-09-18

Bonsai 2 ships a 27B local model in a 5.95 GB download

Prism has released Bonsai 2, a 27B-parameter local model whose smallest main language-weight download is 5.95 GB. The compression result is real, but it is not a 5.95 GB runtime-memory promise or proof that every large model now runs perfectly on a phone.\n\n### Key facts\n\n- PTQ1_0 language weights are a 5.95 GB download; default PQ2_0 is 7.21 GB.\n- Image use adds an optional roughly 0.63 GB Q8 projector on disk.\n- Prism reports 84.78 versus 80.98 for prior Bonsai on the same aggregate.\n- Primary source: Prism's model card.\n\nBonsai 2 stores language weights as ternary negative-one, zero or positive-one values, with group scales and a Hadamard rotation. It is like shipping a book with a compact codebook and requiring a special reader to decode it. The runtime and format guide explain why Prism's fork is needed for the main formats.\n\nDisk size is not VRAM. Running also needs active weights, a key-value cache, activations and image-projector state. Prism names test hardware such as RTX 5090, Apple M5 Max and L4 but does not state a single minimum/recommended VRAM requirement. The verified claim is the 5.95 GB download, not a derived hardware requirement.\n\nPrism's reported 84.78 aggregate versus 80.98 for prior Bonsai is meaningful as a vendor measurement. Its own detailed comparison trails the FP16 base in tool calling and vision, while official documentation supports structured tool calls, MCP and image input. Capability availability is not quality parity.\n\nThe strongest counterargument is deployment friction: bespoke formats matter less if install, hardware variation and multimodal quality disappoint. That is fair. The actual advance is architectural: private text workloads can consider a local 27B model as a 5.95 GB download. Long context, vision, tool use and production reliability still require independent evidence.

Primary source, verified: read the paper →

Ground Truth. (2026, September 18). Bonsai 2 ships a 27B local model in a 5.95 GB download. Ground Truth. https://groundtruth.day/news/bonsai-2-ships-a-27b-local-model-in-a-5-95-gb-download.html

@misc{groundtruth:bonsai-2-ships-a-27b-local-model-in-a-5-95-gb-download,
  title  = {Bonsai 2 ships a 27B local model in a 5.95 GB download},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {sep},
  url    = {https://groundtruth.day/news/bonsai-2-ships-a-27b-local-model-in-a-5-95-gb-download.html}
}

Topics: open-weights · local-ai · quantization · multimodal · tools