Ternary Bonsai 2 27B Requirements: 5.9GB, VRAM & GPU (2026)

Ternary Bonsai 2 27B is a ternary-quantized derivative of Qwen3.8-27B that PrismML released on September 17, 2026, needing just 5.9GB of memory at 1.76 effective bits per weight — it runs on a 16GB-RAM laptop or a single 24GB GPU such as an RTX 4090. That's more than a 9x reduction from the 54GB FP16 reference, yet the official benchmark shows it retains 98.2% of Qwen3.8-27B's (FP16) accuracy.

This article is a requirements at-a-glance guide plus a comparison against the previous release (PrismML Bonsai 27B) and the base model (Qwen3.8-27B).

In July 2026, PrismML released Bonsai 27B, a post-hoc ternary quantization of Qwen3.6-27B (1.71 effective bits, 5.9GB, 94.6% of FP16 retained). Ternary Bonsai 2 27B is version 2 of that line, rebuilt on top of Qwen3.8-27B — a newer base with hybrid linear/full attention, a 262K context window, and native multimodal input.

Beyond the newer base model, PrismML also refined its quantization method. The official announcement reports an "intelligence density" (a performance-per-GB metric) improvement from 0.416 to 0.457, a +9.9% gain, while the footprint stays in essentially the same 5.9GB range as before — same size class, more capable inside.

Weights are quantized to three values, {-1, 0, +1}, with one FP16 scale factor shared across each group of 128 weights. This group-wise scheme preserves more information than a naive 2-bit quantization, landing at an official effective bit-width of 1.76 bits/weight. The model card also documents quantization of the vision tower (the image-input encoder) separately.

PTQ1_0 is closest to the theoretical minimum at 1.75 bpw, but PQ2_0's 2-bit slot packing is sometimes preferred for runtime or hardware compatibility reasons. If you use image input, add the vision tower's roughly 0.63GB (Q8_0) on top of the base file.

Math and Coding land at or slightly above FP16 parity, while Knowledge & Reasoning shows the largest drop. That's a familiar pattern for ternary quantization: procedural reasoning survives better than broad factual recall. Compared against IQ2_XXS (72.59), a similarly low-bit format, Ternary Bonsai 2 27B beats it by more than 12 points while sitting at roughly 82% of its size — a large gap versus naive ultra-low-bit quantization.

Generation speed (TG128) and prompt-processing speed (PP512) are separate metrics — check PP512 if your workload involves long input prompts. Apple Silicon (M5-series) already reaches practical speeds, making a Mac-only offline setup a realistic option.

PrismML's announcement and independent reporting both describe the model as running on "16GB of RAM, or a single 24GB GPU." That guidance covers the model weights (roughly 5.9GB) plus typical runtime overhead — it does not account for the KV cache growth that comes with using the full 262K context window**, which will require additional memory beyond that baseline. No official figures break down memory use by context length, so treat the 16GB tier as a floor for short prompts and plan for more headroom if you run long conversations or documents.

The footprint stays in essentially the same 5.9GB class, but a newer base model plus quantization refinements lift accuracy retention from 94.6% to 98.2%. The core value proposition — a 27B-class model in 5.9GB — hasn't changed; what's inside it got meaningfully better.

- llama.cpp: requires PrismML's own fork with ternary kernels (CUDA/Metal/CPU support). Upstream llama.cpp cannot correctly unpack ternary weights
- MLX (Apple Silicon): prism-ml/Ternary-Bonsai-2-27B-mlx-2bit is available for Mac, iPhone, and iPad
- Ollama: not an official distribution, but community uploads exist. See our Ollama vs LM Studio comparison for tool selection guidance
- API (OpenRouter): available without self-hosting at $0.075/M input tokens and $0.50/M output tokens