Flow - 4x faster than typing
Mini toolsBonsai 27B mlx 1bit
By PrismML
Model family: Qwen
Bonsai 27B (1-bit) is a binary-weight quantization of the 27B-parameter Qwen3.6-27B hybrid-attention language model. Each weight is a single sign bit with FP16 group-wise scaling (group size 128), giving a true 1.125 bits per weight and a ~3.9GB footprint, about 14.2x smaller than FP16. It supports a 262K token context with 4-bit KV-cache quantization, includes an optional 4-bit vision tower for image input, and ships with a DSpark speculative-decoding drafter for faster CUDA serving. It retains 89.5% of FP16 thinking-mode benchmark performance across knowledge, math, coding, instruction following, agentic tool use, and vision tasks, staying closest to full precision on math and coding. It runs via Apple MLX (Python and Swift) on iPhone and Mac, and via CUDA on GPUs, enabling on-device 27B-class reasoning on phones and laptops.
New
Multimodal
Overview
1-bit quantized version of the 27B-parameter Qwen3.6-27B language model, using binary g128 weight representation (1.125 bits per weight) for a ~3.9GB deployed footprint, about 14.2x smaller than FP16. Supports a 262K token context, retains around 89.5% of FP16 accuracy across reasoning, math, coding, and tool-use benchmarks, and runs on-device on phones, laptops, and GPUs via Apple MLX and CUDA.
About PrismML
Industry:
Artificial Intelligence
Company Size:
6
Location:
Pasadena, California, US
Website:
prismml.com
Last updated: July 15, 2026
