01
Overview
Bonsai 2 27B is PrismML's second-generation 27B-class ternary model, released September 17, 2026. It is derived from Qwen3.8 27B and is designed to keep much of a 27B model's reasoning, coding, vision and tool-use capability while making local deployment dramatically smaller.
The headline number is 5.95 GB for PrismML's PTQ1_0 language-model package. That is unusually small for a model with roughly 27 billion parameters, but the important details are in the format, runtime requirements and benchmark caveats.
02
What is Bonsai 2 27B?
PrismML describes Bonsai 2 27B as a 27.8-billion-parameter multimodal language model based on Qwen3.8 27B. The language weights use ternary values {-1, 0, +1} with group-wise scaling rather than a conventional 4-bit or 8-bit quantization scheme.
The official GGUF model card lists a 27.36B-parameter model with a 24.35B language backbone, a 2.54B embedding/LM head component and a roughly 0.46B vision tower. The architecture remains based on Qwen3.8's hybrid-attention design.
PrismML publishes the model under the Apache 2.0 license. The company provides GGUF builds for llama.cpp-style inference and an MLX build for Apple Silicon.
03
How small is it?
The official GGUF release lists two main language-model packages:
- PTQ1_0: approximately 5.95 GB.
- PQ2_0: approximately 7.21 GB.
PrismML compares the smaller package with roughly 54 GB for the full-precision base model and describes it as about 9.3 times smaller than FP16.
Vision is optional rather than baked into that 5.95 GB figure. The GGUF model card lists a separate Q8_0 vision projection package of roughly 0.63 GB for image input.
That distinction matters when estimating the actual storage and memory footprint of a multimodal setup.
04
Context length
PrismML's model card lists a 262K-token context window inherited from Qwen3.8 27B. A large advertised context window does not mean every laptop can practically use the full window: runtime memory also depends on context length, cache implementation, backend, prompt composition and whether vision is enabled.
05
What can Bonsai 2 27B do?
The release targets reasoning, mathematics, coding, instruction following, vision and agentic tool use. PrismML says the ternary model preserves thinking/reasoning behavior despite its low-bit representation.
It is also multimodal. With the separate vision component loaded, PrismML's demo supports photos, screenshots and PDFs as inputs.
These capabilities come from the Qwen3.8-derived architecture plus PrismML's compression approach; Bonsai 2 is not a completely unrelated model trained from scratch.
06
The 98.2% benchmark claim
PrismML reports an aggregate score of 83.9 across a 20-benchmark suite versus 85.4 for its full-precision Qwen3.8 27B reference, which it describes as 98.2% capability retention.
The more detailed GGUF model card reports an 84.78 average across 14 thinking-mode benchmarks and says math and coding remain close to full precision. PrismML also reports up to 143 tokens per second on an NVIDIA GeForce RTX 5090 and roughly 47 tokens per second on an Apple M5 Max for a documented configuration.
These are vendor-reported results. They should not be interpreted as independent benchmark reproduction, and performance on a particular computer can differ substantially.
07
An important llama.cpp compatibility caveat
Bonsai 2's compact formats need runtime support for PrismML's ternary representation and activation transform.
PrismML's Bonsai demo repository currently says the Bonsai 2 GGUF and MLX bands require its llama.cpp fork. Its separate development model card warns that stock llama.cpp can load one testing Q2_0 file yet produce invalid output because the required activation transform is missing.
That makes the runtime version more important than the .gguf extension alone. Do not assume that every GGUF-compatible app can correctly run Bonsai 2 today.
PrismML is working toward upstream support, so this compatibility status should be checked again before installing.
08
GGUF vs MLX
The GGUF route is intended for broader local deployment across CUDA, Metal, CPU and other llama.cpp-supported environments, using PrismML's currently required runtime modifications.
The MLX 2-bit release targets Apple Silicon. Hugging Face currently lists that repository as an 8.6 GB MLX package. The representation and packaging differ from the 5.95 GB PTQ1_0 GGUF, so the two download sizes should not be treated as contradictory measurements of the same file.
09
Is Bonsai 2 27B free?
Yes, the downloadable model is published under Apache 2.0 and PrismML announced it as free to download. Local use still carries the ordinary hardware, electricity and storage costs of inference.
10
Can it run on a laptop?
PrismML explicitly targets laptops, consumer CPUs and edge GPUs, and provides Apple Silicon examples. But there is no single official minimum-RAM or minimum-VRAM number that applies to every context length, backend and modality.
The 5.95 GB model file is not the same thing as total runtime memory. Applications also need memory for runtime state, context/cache, activations and optional vision components. Hardware advice should therefore be based on tested configurations rather than simply adding a small margin to the model file size.
11
Why Bonsai 2 matters
Most local-model optimization has focused on conventional quantization. Bonsai 2 is interesting because PrismML is pushing ternary weights into a capable 27B multimodal model while trying to preserve reasoning and tool-use behavior.
If the approach generalizes and runtime support matures, it could make larger local models practical on hardware that previously required much smaller parameter counts. Independent testing will be important for determining where the quality trade-offs actually appear.
12
FAQ
Is Bonsai 2 27B open source?
The model weights are downloadable under Apache 2.0. "Open-weight" is the most precise description unless referring specifically to source code in PrismML's public repositories.
What model is Bonsai 2 based on?
Qwen3.8 27B.
How large is Bonsai 2 27B?
The primary PTQ1_0 GGUF language model is about 5.95 GB; PrismML also provides a roughly 7.21 GB PQ2_0 package. Vision requires a separate component.
Does Bonsai 2 support images?
Yes. PrismML documents an optional vision tower and its demo supports image inputs.
What is its context window?
The official model card lists 262K tokens inherited from the base model.
Does Bonsai 2 work with stock llama.cpp?
Not reliably for the current compact Bonsai 2 formats. PrismML currently directs users to its fork while upstream support is being developed.
Are the 98.2% benchmark and speed figures independent?
No. They are PrismML's reported results and should remain attributed until independently reproduced.
Sources
Primary and supporting sources
Use the linked documentation to check current pricing, access and product behaviour. The dates above describe this article; they do not guarantee that every vendor claim remains current.