MichelRosselli/ bonsai-27b:Q1_0

138 3 days ago

Full 27B-class reasoning in binary transformer weights.

vision tools thinking 
ollama run MichelRosselli/bonsai-27b:Q1_0

Details

3 days ago

69db21d6b602 · 4.4GB ·

model
qwen35
·
26.9B
·
(!unknown_file_type 40!)
projector
clip
·
461M
·
Q8_0
license
Apache License Version 2.0, January 2004 http://www.apache.org/licenses/ TERMS AND CONDITIONS FOR US
license
This software is copyright 2026-present Prism ML, Inc. It is available under the Apache 2.0 license.
params
{ "min_p": 0, "num_ctx": 262144, "repeat_penalty": 1, "stop": [ "<|im_start|

Readme

Bonsai 27B

This is an unofficial Ollama upload of Prism ML’s 1-bit Bonsai 27B model. It packages the Q1_0 language model with the Q8_0 vision projector and uses Ollama’s native Qwen 3.5 renderer and parser for vision, tool calls, and thinking.

Tags

Tag Size Context Notes
latest 4.4 GB 262,144 Alias of Q1_0
Q1_0 4.4 GB 262,144 Q1_0 language model with Q8_0 vision projector

Usage

ollama run MichelRosselli/bonsai-27b

ollama run MichelRosselli/bonsai-27b:Q1_0

Model

Item Specification
Base model Qwen3.6-27B
Language model 26.9B parameters, Q1_0
Vision projector 460.73M parameters, Q8_0
Context window 262,144 tokens
Ollama version 0.32.5 or newer
Capabilities Text, vision, tools, thinking

Prism ML describes Bonsai 27B as a binary-weight model with a true 1.125 bits per language-model weight and an approximately 3.9 GB language-model footprint. The published Ollama package is 4.4 GB because it also includes the optional Q8_0 vision projector.

Generation Settings

Parameter Value
temperature 0.7
top_k 20
top_p 0.95
min_p 0.0
repeat_penalty 1.0
num_ctx 262,144

num_predict remains request-specific. Ollama does not expose a max_p Modelfile parameter.

Notes

  • This upload uses RENDERER qwen3.5 and PARSER qwen3.5; it does not replace the native Ollama renderer/parser with a custom template.
  • The Q8_0 multimodal projector is included and image input was runtime-tested.
  • DSpark is not included because the tested Ollama releases do not support its architecture or expose it through a DRAFT Modelfile directive.
  • The F16 variant is intentionally not included; Q1_0 is the lightest compatible language-model variant in the source repository.
  • Text, thinking, structured tool calls, long histories, and front-loaded image prompts were compared with the canonical Hugging Face template.
  • Ollama cannot exactly represent interleaved text/image ordering, video placeholders, add_vision_id, or preserve_thinking through the native renderer API.
  • With think=false, Ollama omits reasoning from the current tool-call history where the Hugging Face template preserves it.
  • Array and object tool arguments differ byte-for-byte only in JSON separator whitespace; schema-aware response parsing reconstructs their types.
  • A literal complete <tool_call> example in normal answer text may be parsed as a real tool call. An unterminated tool-call block may be discarded by the native parser.

Attribution

All credit for the original model belongs to Prism ML.

Created using Bonsai by Prism ML.

Original resources:

License: Apache 2.0.