Short answer: yes. Bonsai 27B is an official Apache 2.0 local release with a 3.9GB 1-bit build and a 7.2GB deployed ternary build. For most desktop users, the ternary variant is the better quality choice on 16GB or more. The 1-bit build is the extreme footprint option. Both currently rely on PrismML's custom low-bit kernels, so this is not yet a normal drop-in quant for every GGUF app.
What Bonsai 27B actually is
Bonsai 27B is a PrismML compression of Qwen 3.6 27B. It keeps the original hybrid-attention architecture, multimodal vision tower, reasoning behavior and agentic tool-use target while replacing conventional high-precision weights with native binary or ternary representations.
This is more than a renamed community quant. PrismML publishes the weights, custom GGUF kernels, MLX variants, benchmarks and methodology under Apache 2.0. The release therefore qualifies for the LocalClaw catalogue as a distinct local model family with two deliberate operating points.
1-bit versus ternary: which file should you choose?
| Variant | Deployed weights | Vendor benchmark average | Best fit |
|---|---|---|---|
| 1-bit Q1_0_g128 | ~3.9GB | 76.11, 89.5% of FP16 | Smallest footprint, phones and low-memory experiments |
| Ternary Q2_0_g128 | ~7.2GB | 80.49, 94.6% of FP16 | Laptops and desktops where quality matters |
| Qwen 3.6 27B Q4_K_XL | ~17.6GB | 84.99, 99.9% of FP16 | Higher quality when memory is available |
The distinction between ideal and deployed size is important. PrismML describes ternary Bonsai as 1.71 effective bits per weight, which is 5.9GB at the information-theoretic representation size. Today's kernels store the values in 2-bit slots, producing an actual deployed language-model file of about 7.2GB. LocalClaw uses the deployed figure.
Can Bonsai 27B run on your machine?
Weight size is not peak memory. The runtime also needs activations, buffers and a KV cache. PrismML measures the 1-bit build at 5.2GB peak for a 4K context and the ternary build at 8.4GB. At 100K context without KV-cache compression, those figures rise to 11.6GB and 14.7GB.
| Hardware | Recommended build | LocalClaw verdict |
|---|---|---|
| 8GB unified memory / RAM | 1-bit, short context | Technically possible, but tight once the OS and other apps are included. |
| 16GB laptop or desktop | Ternary for quality, 1-bit for headroom | The practical consumer starting point. |
| 24GB to 32GB | Ternary with longer context | Comfortable daily local use and multimodal experiments. |
| RTX 4090 / 5090 | Ternary or 1-bit CUDA | Fast single-GPU target using PrismML's CUDA kernels. |
How fast is it?
PrismML reports 44.2 tokens per second for the 1-bit build and about 26 tokens per second for ternary on an Apple M5 Pro. On an M4 Pro, the published 1-bit result is 26 tokens per second. The launch post reports higher best-case figures on an M5 Max and RTX 5090.
These are vendor measurements using custom low-bit kernels, not universal LM Studio numbers. They show that the format can be fast when the runtime is tuned for it. They do not guarantee the same throughput in every app, prompt length or hardware configuration.
The runtime caveat most headlines miss
The files are GGUF, but the Q1_0_g128 and Q2_0_g128 formats use custom kernels for Qwen 3.6's hybrid-attention architecture. PrismML's official quickstart currently points to its own llama.cpp fork for Metal, CUDA and CPU, plus dedicated MLX and MLX Swift releases.
That means you should not assume every stock GGUF application supports Bonsai correctly today. Mainstream integration may improve quickly. Until then, follow the official PrismML runtime instructions. LocalClaw lists the official GGUF repository, but labels the custom-runtime requirement clearly.
How does it compare with current local favorites?
| Model | Why choose it | Why not |
|---|---|---|
| Bonsai 27B ternary | Unusually high 27B capability per GB, vision and 262K context | Specialized runtime and some quality loss |
| Qwen 3.6 27B Q4 | Near full-precision quality and broader runtime familiarity | Roughly 17.6GB just for weights |
| Gemma 4 12B | Strong multimodal balance and conventional local formats | Smaller model class and different capability target |
| GPT-OSS 20B | Strong reasoning and tool use with Apache 2.0 | Less aggressive memory density |
What about the phone claim?
PrismML publishes an MLX Swift build and reports that 1-bit Bonsai runs on an iPhone 17 Pro Max. That is technically significant. It is also a specialized deployment, not the experience of downloading a random GGUF into a mobile chat app. One launch demonstration explicitly uses cached and prefilled image context, so it should not be interpreted as full image processing from a cold start in every scene.
LocalClaw verdict
Bonsai 27B is one of the most interesting local AI releases of 2026 because it moves a modern 27B multimodal reasoning model into a memory tier normally occupied by much smaller models. The quality-oriented ternary build is the sensible default. The 1-bit build is the technical showcase and the right choice when footprint matters more than maximum quality.
Use it if you enjoy testing new runtimes, want 27B-class behavior on 16GB hardware, or need a private agent with long context. Wait if you need a zero-friction LM Studio workflow today. This is an important release, but the runtime ecosystem is still catching up to the weights.