Leaderboard / PrismML
Ternary Bonsai 2 27B
PrismML · Bonsai · released 2026-09-17 · data checked 2026-11-03
Ternary Bonsai 2 27B is a 27.4B-parameter dense model from PrismML (United States), released 2026-09-17 under the Apache-2.0 license. It scores 74.6 on the Small LM Index, #3 of 34 medium models. It needs about 18.2 GB of memory at 4-bit with an 8K-token context.
Qwen3.8 27B compressed to ternary weights (about 1.7 bits each): the file is about 6 GB instead of 54 GB, and PrismML reports it keeps 98% of the original average score. Scores are PrismML's own runs (EvalScope, thinking mode). The smallest files need PrismML's llama.cpp fork; the Q2_0 file runs on stock llama.cpp b11556 or newer.
Benchmarks
Each bar shows this model's score. The tick marks the best medium model on that benchmark. Sources: 8 reported by PrismML. How we label sources.
Not reported: estimated
What the index model expects from this model's other scores. Not a measurement.
Specifications
- Lab
- PrismML (US · United States)
- Based on
- Qwen3.8 27B (fine-tune or distillation)
- Architecture
- Dense transformer
- Total parameters
- 27.4B
- Active parameters
- All (dense)
- Size class
- Medium
- Context window
- 262K tokens
- Input modalities
- text, image
- Reasoning
- Reasoning (thinking)
- License
- Apache-2.0
- Openness
- Open weights
- Memory to run
- ~18.2 GB at 4-bit, 8K context (more sizes)
Compare with
- Ternary Bonsai 2 27B vs Qwen3.8 27B (27B, index 77.0)
- Ternary Bonsai 2 27B vs Gemma 3 27B (27.4B, index 32.7)
- Ternary Bonsai 2 27B vs Qwen3.6 27B (27B, index 74.0)
- Ternary Bonsai 2 27B vs Qwen3.5 27B (27B, index 70.6)
Ranked in
- #2 in Best small open models for coding
- #2 in Best small open models for math
- #3 in Best LLMs for a 24 GB GPU (RTX 4090, RTX 3090)
- #2 in Best small open vision-language models
- #3 in Best open models with an Apache 2.0 or MIT license
- #1 in Best American open models under 100B
Sources
- https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-gguf
- https://openrouter.ai/prism-ml/ternary-bonsai-2-27b
Something wrong here? Report it.
Run it locally
Uses prism-ml/Ternary-Bonsai-2-27B-gguf (Q2_0). Swap the tag for another size: F16, PQ2_0, PTQ1_0, Q2_0.
llama-server -hf prism-ml/Ternary-Bonsai-2-27B-gguf:Q2_0This model ships in its own Q2_0 format; Ollama and LM Studio may not load it yet.
Memory needed: weights, context cache and runtime buffers.
| Quantization | 4K context | 32K context | 256K context |
|---|---|---|---|
| 16-bit F16 | 56 GB | 58 GB | 73 GB |
| Q2_0 (as released) Q2_0 | 9.0 GB | 11 GB | 26 GB |