Bonsai 27B
#14 in Reasoning ModelsUnknown · 27b · 2× · last seen Aug 01, 2026
Bonsai 27B is an open-source Reasoning model with 27.8 billion parameters released by PrismML on July 14, 2026, created through extreme low-bit compression (1-bit or ternary) of the base model Qwen3.6 27B. It comes in two variants – a 3.9 GB 1-bit version for smartphones and a 5.9 GB ternary version for laptops – and supports text, vision, tool-calling, and agentic workflows with a context window of 262,144 Token. According to the manufacturer, the ternary variant retains approximately 94.6–95% and the 1-bit variant retains approximately 89.5–90% of the Benchmark performance of the uncompressed base model. The model is available free of charge under the Apache 2.0 license as weights on Hugging Face as well as through free third-party API endpoints.
Features
| Key Benchmark (%) | 80.49 average across 15 benchmarks (Ternary, thinking mode) = 94.6–95% of FP16 baseline; 1-bit variant 76.11 (89.5–90%) |
| Context Window (Tokens) | 262,144 tokens |
| License | Apache 2.0 (open weights on Hugging Face) |
| Multimodality | Yes – text, vision (4-bit vision tower), tool calling, agentic workflows |
| Platform | On-device: iPhone (17 Pro/Pro Max), laptops/PCs (llama.cpp, MLX, RTX 5090, M5 Pro); also cloud APIs (Puter, Together AI) |
| Price per 1M Tokens | $0 (free, e.g. via Puter and Together AI API, and open weights) |
| Release Date | July 14, 2026 |