Bonsai 27B is a new open-weights multimodal model that PrismML compressed down to as little as 3.9 GB, small enough to run entirely on a phone or laptop with no cloud connection. Released July 14 under the Apache 2.0 license, it is the first 27B-class model designed to run on-device, and it keeps most of the quality of the full-size model it is built on.

Try It: Run a 27B Model Locally Today

You can download the weights right now from the Bonsai 27B collection on Hugging Face. Pick the ternary build (5.9 GB) if you have headroom, or the 1-bit build (3.9 GB) for the tightest phone and laptop budgets. If you just want to test it before downloading anything, PrismML ships an in-browser WebGPU demo that runs the model client-side. Smaller 8B, 4B, and 1.7B variants are also live in the collection for lower-end hardware.

Why It Matters for Creators

On-device models mean your prompts, images, and drafts never leave your machine, and generation costs nothing per token once the weights are local. Bonsai is built on Qwen3.6 27B and, according to PrismML's published benchmarks, the ternary version retains about 95 percent of the base model's overall score while the 1-bit version holds roughly 90 percent. That puts genuine 27B-class reasoning, coding, and vision into the same local-first workflows creators already build around tools like Ollama.

Key Details

Sizes: ternary (5.9 GB, 1.71 bits per weight) and 1-bit (3.9 GB, 1.125 bits per weight).

Base and context: Qwen3.6 27B, 262K-token context, text plus vision.

Speed: up to 163 tokens per second on an RTX 5090 and 87 tokens per second on an M5 Max for the 1-bit build.

License: Apache 2.0, with the compression method documented in the project whitepaper. The model is also served through the Together AI API for anyone who wants it hosted.

What to Do Next

Match the build to your device: the 1-bit model for phones and thin laptops, the ternary model for a desktop GPU, and one of the smaller variants if memory is tight. Load it into your existing local runner, point your usual coding or vision prompts at it, and compare the output against whatever cloud model you currently pay for. If it holds up on your workload, you have just moved a frontier-class model fully in-house.