Ternary Bonsai 2 27B

Use it for local reasoning, coding, multimodal understanding, tool use, and agentic workflows on consumer-grade CPUs and edge GPUs.

Availability

Free download starting September 17, 2026

Pricing model

Free

Deployment

Optimized for consumer-grade CPUs and edge GPUs

Privacy

Not established

Licence

Apache 2.0

• Bonsai 2 needs an activation transform that currently only exists in PrismML’s llama.cpp fork. ^{[1] [4]}

  Bonsai 2 needs an activation transform that currently only exists in our

• The Q2_0 packing is kept separate until upstream support lands. ^[1]

  The `Q2_0` packing of Bonsai 2 27B, kept separate until upstream support lands.

• The build will move to the main repo once upstream support lands.
• The build is intended for testing, kernel work, and upstreaming.
• Stock llama.cpp may load the file and output gibberish with no warning. ^[4]
• The model supports multimodal text-and-image input and a 262K-token context window. ^[2]

  It supports a 262K-token context window , multimodal text-and-image input, and is released under the Apache 2.0 license.

• The model uses ternary weights with FP16 group-wise scaling and has a total model footprint of 5.9GB. ^[2]

  Ternary Bonsai 2 27B uses ternary {−1, 0, +1} weights with FP16 group-wise scaling, for 1.76 effective bits per weight and a total model footprint of 5.9GB .

• The model runs on NVIDIA GPUs via CUDA and Apple devices via MLX. ^[2]

  Bonsai 2 27B runs on NVIDIA GPUs via CUDA and on Apple devices (Mac, iPhone, iPad) via MLX, through custom low-bit kernels.

• The model is intended for local assistants, multimodal agents, private knowledge workflows, computer-use applications, and long-running agentic tasks. ^[3]

  It is intended for local assistants, multimodal agents, private knowledge workflows, computer-use applications, and long-running agentic tasks that have traditionally depended heavily on cloud inference.

• Users can download the model for free starting September 17, 2026. ^[3]

  Developers, researchers, and other users can download the Ternary Bonsai 2 27B models under the Apache 2.0 license for free starting September 17, 2026.

• The ternary model is 5.9 GB and reduces memory footprint by more than 9x compared with its full-precision counterpart. ^[3]

  At just 5.9 GB, Ternary Bonsai 2 27B reduces memory footprint by more than 9x compared with its full-precision counterpart.

• It retained 98.2% of Qwen3.8 27B’s benchmark performance. ^[3]

  However, while the original Ternary Bonsai 27B retained 95% of the aggregate benchmark performance of its full precision counterpart, the new Ternary Bonsai 2 27B retained over 98%.

• It achieved an aggregate score of 83.9 across a 20-benchmark suite. ^[3]

  Across a 20-benchmark suite spanning reasoning, math, coding, instruction following, vision, and agentic tool use, Ternary Bonsai 2 27B achieves an aggregate score of 83.9 , retaining 98.2% of Qwen3.8 27B’s performance of 85.4 .

• The model is a 27.8-billion-parameter multimodal language model. ^[3]

  The Ternary Bonsai 2 27B model is a 27.8-billion parameter multimodal language model.