Prism ML Releases Ternary-Bonsai-2-27B-gguf: A 1.72 bpw Ternary Model

September 18, 2026

At a Glance

Values determined by this site’s code at collection time. Dates are JST.

Overview

Prism ML has released the GGUF version of its ternary language model based on Qwen3.8-27B, named prism-ml/Ternary-Bonsai-2-27B-gguf. This model achieves an extremely compact weight representation of an effective 1.72 bits/weight, achieving approximately 9.3x reduction in size compared to FP16 while maintaining 98.2% of the intelligence score of the FP16 model in thinking mode.

By combining proprietary rotated basis transformation and ternary representation, it avoids the degradation of reasoning and thinking capabilities that typically occurs with conventional ultra-low-bit quantization. It is developed with the aim of running advanced 27B-class reasoning and agent operations locally on standard laptops or single GPU environments.

Specifications

• Parameter Count: 27B class (26.2M parameters including recurrent state paths and normalization weights are preserved in high precision)
• Architecture: Hybrid attention structure of Qwen3.8-27B (approximately 75% linear attention Gated DeltaNet, Gated Attention)
• Context Length: 262,144 tokens (262K)

Performance

According to measurements published by Prism ML, Bonsai 2 27B shows remarkable performance retention compared to traditional sub-4-bit quantization (such as IQ2_XXS). A comparison of the key evaluation results listed in the model card is as follows:

The breakdown by skill category (compared to the original FP16 model) is as follows:

Detailed results across all 14 benchmarks are as follows:

According to the published measurement data in the model card, Bonsai 2 27B achieves an average score of 84.78 across 14 benchmarks despite its small size of 1.72 bpw (5.9 GB footprint). It significantly outperforms IQ2_XXS (2.8 bpw, 72.59 points), a conventional 2-bit equivalent method, and approaches the accuracy of the 4-bit quantized model UD-Q4_K_XL (85.18 points).

In particular, it maintains high accuracy comparable to or exceeding full precision (FP16) in mathematics (such as GSM8K for grade-school word problems and MATH-500 for high school/competition-level math problems) and coding (such as HumanEval+ for implementing short functions according to specifications, MBPP+ for beginner-to-intermediate tasks, and LiveCodeBench for time-segmented fresh programming challenges). It also scores an excellent 91.31 on IFEval, which measures adherence to format constraints such as “in 300 words or less." On the other hand, a slight drop in accuracy compared to the full-precision version is observed in visual domains such as MMMU-Pro and OCR Bench v2, which evaluate knowledge and multimodal understanding, showing the model’s strengths and weaknesses.

Strengths and Use Cases

Bonsai 2 27B excels at advanced mathematical reasoning in Thinking Mode, code generation, and multi-step tool calling (agent operations). Equipped with a wide context length of 262K, it is also well-suited for analyzing massive codebases and reading long-form documents.

Additionally, it can support multimodal tasks when combined with a vision encoder (a separate Q8_0 mmproj file). Due to its small model size, intended use cases include local execution on standard laptops, privacy-focused offline environments, and low-latency serving with a single GPU.

How It Differs from Similar Models

The differences from Qwen3.8-27B derivative models previously introduced on this site are as follows:

Salience-27B-R6 GGUF Quantized Version Released, a Thinking-Optimized 27B Multimodal Model is a derivative model that focused on “Reasoning economy," which shortens the thinking process to reduce token consumption. In contrast, Bonsai 2 27B aims to drastically minimize the model structure itself down to 5.9 GB using proprietary ternary quantization technology while preserving full-precision reasoning accuracy as much as possible.

Qwen3.8-27B Speed-Optimized Fine-Tune “Signal-3.8-27B-GGUF" Released is a fine-tuned model that reduces the number of generation and thinking tokens to shorten response times. Bonsai 2 27B is differentiated by its direction of enabling on-device operation through innovations in weight quantization representation (ternary format) rather than reducing token counts.

Hardware Requirements

Estimated requirements (calculated by Local Model Watch) — 27.8B parameters (taken from the base model Qwen/Qwen3.8-27B)

Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model’s authors. Compare with other models in our VRAM quick reference. What the quantization names mean: glossary.

Recent Models in the Same Size Class

Models with 15–40B parameters that Local Model Watch covered recently, listed by code from our article log for comparison. VRAM tiers are this site’s estimates; licenses are as stated on the model cards.

How to Get It

It is distributed in GGUF format and can be obtained from the Hugging Face repository prism-ml/Ternary-Bonsai-2-27B-gguf. No license agreement (gated) is required.

Because this model uses proprietary ternary hybrid attention kernels (PTQ1_0 and PQ2_0 packing), it will not work with standard llama.cpp. You must use the fork of llama.cpp provided by Prism ML or build it from source.

The model can be downloaded using huggingface-cli (the hf command).

hf download prism-ml/Ternary-Bonsai-2-27B-gguf Ternary-Bonsai-2-27B-PQ2_0.gguf --local-dir.

An example execution using the Prism ML version of llama.cpp (llama-cli) is as follows:

./bin/llama-cli -m Ternary-Bonsai-2-27B-PQ2_0.gguf \
    -ngl 99 -fa on -c 32768 \
    --temp 1.0 --top-p 0.95 --top-k 20 \
    -p "Explain quantum computing in simple terms." -n 256

Note that a separate repository for MLX (Ternary-Bonsai-2-27B-mlx-2bit) is also available for Apple Silicon.

Quantized and Converted Variants

In addition, 3 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files.

Related Articles

• Salience-27B-R6 GGUF Released by bartowski
• Signal-3.8-27B-GGUF: Faster and More Token-Efficient

Sources

• prism-ml/Ternary-Bonsai-2-27B-gguf
• Qwen/Qwen3.8-27B
• Bonsai 2 27B Whitepaper
• Bonsai Demo & Examples Repository
• llama.cpp (PrismML Fork)