PrismML's Ternary Bonsai 2: a 27B-class model in 5.95 GB
PrismML released Ternary Bonsai 2: Qwen models stored with three-value weights, from a 27B model in 5.95 GB to a 2B in 0.43 GB, with image input, a 262K context and Apache 2.0.
PrismML, a company in Pasadena, California, released Ternary Bonsai 2 in two steps: the 27B model on 17 September 2026, and the 9B, 4B and 2B models on 15 October. Each is an Alibaba Qwen model whose weights are stored with just three values, which makes the 27B file 5.95 GB and the 2B file 0.43 GB.
What is new
The models are not trained from scratch. Each is derived from a Qwen model with the architecture unchanged: Qwen3.8-27B for the largest and Qwen3.5 for the 9B, 4B and 2B. Every language-model weight, including the embeddings and the output layer, is stored as −1, 0 or +1, with one 16-bit scale for each group of 128 weights. That comes to about 1.7 bits per weight instead of 16, an extreme form of quantization known as ternary weights. All four sizes keep the base model’s 262K-token context, accept images, call tools and think before they answer by default.
PrismML reports that on its own 14-benchmark suite the 27B model averages 84.78, which is 98.2% of the 86.32 scored by the 16-bit Qwen3.8-27B, while a conventional 2-bit build of the same model averages 72.59. The losses are in knowledge and reasoning (79.86 against 85.55) and in vision (66.19 against 71.36). The small sizes lose more: the 2B scores 39.6 on GPQA Diamond where Qwen3.5-2B scores 58.48. These are PrismML’s own results, and we have not reproduced them.
Running it
The 27B language model is 5.95 GB in the densest GGUF packing, called PTQ1_0, against about 54 GB at 16-bit, and a 7.21 GB packing called PQ2_0 decodes faster on some GPUs. The 9B, 4B and 2B files are 2.00, 0.95 and 0.43 GB. Image input needs a separate vision-projector file of 0.36 to 0.63 GB. The two compact packings need PrismML’s llama.cpp fork, or its MLX build on Apple silicon. A third file type, Q2_0, is larger (7.63 GB for the 27B) and runs on recent builds of upstream llama.cpp, but older builds load it without a warning and produce garbage. PrismML reports about 130 tokens per second for the 27B on an RTX 5090 and about 28 on an Apple M5 Pro laptop. The default reasoning effort is xhigh, so give the model a generous output limit.
License
All four sizes use Apache 2.0, the same license as their Qwen base models, so commercial use is allowed. PrismML’s first-generation Bonsai models, in 1-bit and 1.58-bit versions, remain on Hugging Face but are superseded.
Why it matters
A 27B-class model with image input that fits in about 6 GB is within reach of an ordinary laptop, and the 4B and 2B sizes at 0.95 and 0.43 GB suit phones, tablets and always-on background tasks. The price is a quality drop that grows at small sizes, and a dependence on PrismML’s fork for the best packings.
For other small models see small models for laptops, which also lists NVIDIA’s Dory 1.0. Quantization explained shows how low-bit formats trade size for quality, and Qwen3.5 is the family behind the three smaller sizes.
Sources
Models in this story
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Ternary Bonsai 2 PrismML | 2.21B – 27.4B | 256K | Apache 2.0 (Permissive) | 0.473 GB Laptop or phone | |
| Qwen3.8-27B Alibaba Qwen | 27.8B | 256K | Apache 2.0 (Permissive) | 19 GB Gaming GPU | |
| Qwen3.5 Alibaba Qwen | 873M – 403B 17B active | 256K | Apache 2.0 (Permissive) | 0.602 GB Laptop or phone | |
| Dory 1.0 NVIDIA | 4B | 256K | OpenMDW-1.1 (Permissive) | 2.76 GB Laptop or phone |