PrismML releases Bonsai 27B LLM in 4GB VRAM

PrismML has released Bonsai 27B, a 27-billion-parameter language model built from Qwen3.6-27B and compressed using a 1-bit quantization format that fits its weights into roughly 3.9GB of VRAM. Independent testing confirms the model works well on math and coding benchmarks while retaining about 89.5 percent of the original score, though instruction following and agentic tool calling suffer, and running it comfortably requires an 8GB or 12GB GPU to account for KV cache memory. Open article →