LocalClaw/LLM/Bonsai 27B
Open-weight local LLM

Bonsai 27B

PrismML low-bit model derived from Qwen 3.6 27B. Official Apache 2.0 ternary (7.2GB deployed) and 1-bit (3.9GB) builds retain multimodal, reasoning and agentic capabilities through custom GGUF and MLX runtimes.

16 GB sweet spot 16 GB RAM Ternary Q2_0_g128 Private local coding assistant
Open official runtime Compare all models
Parameters
27.3B (ternary / 1-bit)
Minimum RAM
16 GB
Model size
7.2 GB
Quantization
Ternary Q2_0_g128

Can Bonsai 27B run locally?

Bonsai 27B is a practical pick for 16 GB machines, especially with Ternary Q2_0_g128 quantization.

Use the official PrismML llama.cpp / MLX setup. The current low-bit files are not a stock LM Studio install.

Model sourceprism-ml/Ternary-Bonsai-27B-gguf
chatcodereasoningvisionagenticmultimodaledgespeed

Install path

01
Check RAM fitMinimum 16 GB RAM. Start with the Ternary Q2_0_g128 quant.
02
Load the modelFollow the official PrismML llama.cpp / MLX instructions. Stock LM Studio support is not confirmed.
03
Control locallyUse LocalClaw to manage models, agents, chat, channels and scheduled OpenClaw work.

Strengths

  • Official Apache 2.0 GGUF and MLX releases from PrismML
  • Quality-oriented ternary build deploys in about 7.2GB
  • Experimental 1-bit build deploys in about 3.9GB
  • 262K context with optional 4-bit KV-cache compression
  • Preserves vision, reasoning, coding and agentic tool-use capabilities
  • PrismML reports 26 tok/s for ternary and 44 tok/s for 1-bit on an Apple M5 Pro

Limitations

  • The low-bit GGUF builds currently depend on PrismML custom llama.cpp kernels
  • Ternary quality remains below the full-precision Qwen 3.6 27B reference
  • The most impressive benchmark and throughput numbers are vendor-reported
  • Long context still increases peak memory well beyond the model file size
  • Phone demonstrations use specialized MLX Swift deployment and one published demo uses cached, prefilled image context

Best use cases

  • Private local coding assistant
  • On-device reasoning
  • Local tool-use agents
  • Multimodal document analysis
  • Long-context research on 16GB+ laptops
  • Low-memory local AI experiments

Capability profile

speed
9
quality
8
coding
9
reasoning
9

Technical notes

Developer
PrismML
License
Apache 2.0
Context window
262,144 tokens
Architecture
Qwen 3.6 27B-derived hybrid-attention multimodal transformer compressed into native ternary Q2_0_g128 and binary Q1_0_g128 weights.

This model fits these next steps

Hardware fit is based on LocalClaw's RAM tier, model size and quantization metadata. Always leave memory headroom for your OS and runtime.

Starter desktopMac mini M4 16GBPortable fitMacBook Air M4 16GBBest 16GB models16GB RAM guide

Similar models to compare

Qwen 3.6 (27B) 27BQwen 3.6 35B-A3B 35B (3B active, MoE)Gemma 4 12B 12BGPT-OSS (20B) 20B

Where to go next