Ternary Bonsai 2 27B

// Overview

Ternary Bonsai 2 27B is PrismML's compressed republication of Qwen3.8 27B, announced on 17 September 2026. Every weight is stored as one of three values — −1, 0 or +1 — with group-wise FP16 scaling, which PrismML reports as 1.76 effective bits per weight and a total model footprint of 5.9 GB, "more than 9x smaller" than the full-precision original.

The pitch is that almost none of the capability goes with the bits. PrismML reports an overall score of 83.9 against 85.4 for full-precision Qwen3.8 27B — "retaining 98.2% of aggregate benchmark performance" — and publishes the per-capability breakdown that is transcribed below, including the older Qwen3.6 27B as a second reference column. Instruction following is the one row where the compressed model comes out ahead of its own base.

The model card describes the architecture as unchanged from the base, a hybrid-attention causal language model, and keeps the base model's 262K-token context, "kept practical on-device by the predominantly linear-attention backbone". Input is multimodal text and image; the vision tower is a separate optional 0.63 GB Q8_0 file.

Two GGUF packings ship: PTQ1_0 packs trits densely at 1.75 bits per weight for 5.95 GB, and PQ2_0 stores each trit in a 2-bit slot at 2.13 bits per weight for 7.21 GB. There is no hosted API — the weights are the product, published on Hugging Face under Apache 2.0, and the model card is explicit that "Stock llama.cpp will not run these files": running them needs PrismML's own llama.cpp fork, which supports CUDA, Metal and CPU backends. PrismML reports up to 143 tokens per second on an NVIDIA GeForce RTX 5090 and 46.8 tokens per second on an M5 Max.

// Benchmarks

This model's scores

1. Overall (PrismML aggregate)83.9%
2. Math (AIME 2026/2025, GSM8K, MATH-500)96.57%
3. Knowledge & Reasoning (MMLU-Redux, GPQA Diamond, AA-LCR)83.95%
4. Instruction Following (IFBench, IFEval)82.66%
5. Coding (HumanEval+, LiveCodeBench v6, MBPP+, BigCodeBench)81.58%
6. Vision (CharXiv, A-OKVQA, OmniDocBench v1.6, RealWorldQA, OCRBench v2)78.59%
7. Agentic & Tool Calling (τ²-bench, BFCLv3)77.57%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• A 27B-class multimodal model in a 5.9 GB footprint, small enough for consumer GPUs and laptops
• 98.2% of full-precision Qwen3.8 27B's aggregate benchmark score, per PrismML's own comparison
• Scores above its full-precision base on instruction following (82.66 vs 81.25)
• 262K-token context inherited from the base model
• Apache 2.0 weights with both GGUF (CUDA/Metal/CPU) and MLX 2-bit packings

// Best for

• Running a 27B-class multimodal model locally on a single consumer GPU or an Apple-silicon laptop
• On-device or air-gapped deployments where the full-precision checkpoint will not fit in memory
• Cutting inference energy and memory cost when the workload tolerates a ~1.5-point aggregate drop
• Benchmarking how far ternary quantisation can be pushed on a current open-weight model