Prism ML ships a 27B-class model in a 5.9GB footprint

Prism ML has released the second generation of its Bonsai series. Ternary Bonsai 2 27B, based on Qwen3.8 27B, stores weights in a ternary {−1, 0, +1} representation with FP16 group-wise scaling, giving 1.76 effective bits per weight and a total model footprint of 5.9GB. The low-bit representation is applied end to end across the language model.

The model supports a 262K-token context window, takes text and image input, and is released under the Apache 2.0 license. According to the company, it is more than 9x smaller than its full-precision counterpart while retaining 98.2% of aggregate benchmark performance, scoring 83.9 on a suite covering reasoning, math, coding, instruction following, vision and agentic tool use. The previous Bonsai 27B generation retained about 95%.

Throughput and energy

Prism ML reports up to 143 tokens per second on an NVIDIA GeForce RTX 5090 and 46.8 tokens per second on M5 Max. On an RTX 4090 the model consumes 0.714 mWh per token, which the company says is 40% more energy-efficient than an 8B model running at full precision.

The company argues low-bit models change the economics of AI deployment beyond local inference: larger models fit into the same memory envelope, and the same hardware can serve more users. Suggested workloads include coding-agent loops, computer-use workflows, private document analysis, multimodal debugging and hybrid setups that keep sensitive tasks local and escalate selectively to the cloud.

Weights are available today. The model runs on NVIDIA GPUs via CUDA and on Apple devices (Mac, iPhone, iPad) via MLX, using custom low-bit kernels. Details of the compression and evaluation process are in the company's whitepaper. Prism ML came out of a team of Caltech researchers and was founded with support from Khosla Ventures, Cerberus and Google, with continuing support from Samsung.

Full announcement: Prism ML.