Ternary Bonsai 2 27B uses ternary weights

Bonsai 2 27B is based on Qwen3.8 27B and has 27.36 billion total parameters.[1][2] Its components include a 24.35 billion-parameter language backbone, a 2.54 billion-parameter embedding and language-model head, and a 0.46 billion-parameter vision tower.[2] PrismML says the model’s ternary weights use only −1, 0 and +1, with an effective representation of 1.76 bits per weight and a total footprint of 5.9GB.[1] The company describes the compressed model as more than nine times smaller than its full-precision counterpart.[1]

PrismML reports an aggregate benchmark score of 83.9, which it says represents 98.2% of Qwen3.8 27B’s aggregate performance.[1] It says benchmark-performance retention improved from 95% in its previous Ternary Bonsai 27B release to more than 98% with Bonsai 2.[1] These are company-reported results; the evidence provided does not specify the benchmarks or their individual scores. PrismML also reports throughput of up to 143 tokens per second on an NVIDIA GeForce RTX 5090 and 46.8 tokens per second on an M5 Max.[1]

The model accepts both text and images and generates text, according to its documentation.[2] PrismML says it can call tools natively and reason before answering.[2] The company says it runs on NVIDIA GPUs through CUDA and on Apple devices through MLX.[1] Its weights are available under the Apache 2.0 license.[1] The evidence does not name a phone model or say that Bonsai 2 has been tested on phones, so the release’s small footprint should not be treated as confirmation of phone compatibility.[1]

PrismML says it released its first Bonsai 27B models two months before this release.[1] TechCrunch reported that the company said its first Bonsai model had been downloaded more than 11 million times, while its smaller models had accumulated another 2.6 million downloads.[3] TechCrunch also reported that PrismML raised a $22.25 million seed round and is backed by Khosla Ventures, Cerberus Capital and Caltech.[3] PrismML says Bonsai 2 uses 0.714 milliwatt-hours per token on an RTX 4090, making it 40% more energy-efficient than an 8B full-precision model.[1]