Pasadena, Calif. – September 19, 2026 -- PrismML has released Bonsai 2 27B, a ternary-compressed version of a 27-billion-parameter multimodal model that runs at just 5.9 GB while retaining 98.2% of its full-precision counterpart's benchmark performance.
Ternary compression shrinks memory footprint by more than 9x
Built on Qwen3.8 27B, Bonsai 2 27B reduces the model's memory requirement to 5.9 GB, a reduction of over nine times compared with the full-precision original. The model contains 27.8 billion parameters and was trained using Google v5 TPUs.
Quality gap narrows sharply from PrismML's first Bonsai release
PrismML's original Ternary Bonsai 27B, launched in July, retained 95% of its full-precision counterpart's aggregate benchmark performance while achieving the same 9x compression. Bonsai 2 27B pushes that retention above 98%, narrowing the accuracy gap between compressed and full-precision models within a single generation cycle.
Model scores 83.9 across a 20-benchmark suite spanning reasoning to agentic tasks
Across benchmarks covering reasoning, math, coding, instruction following, vision, and agentic tool use, Bonsai 2 27B posts an aggregate score of 83.9 against Qwen3.8 27B's 85.4. On an NVIDIA GeForce RTX 5090, the model runs at up to 143 tokens per second.
PrismML positions the release for on-device agentic workloads
The company built Bonsai 2 27B for local assistants, multimodal agents, private knowledge workflows, computer-use applications, and long-running agentic tasks that have typically relied on cloud inference. The model is optimized for low-latency inference on consumer-grade CPUs and edge GPUs.
"Bonsai 27B proved that powerful models do not have to be confined to cloud infrastructure," said Babak Hassibi, Founder and CEO of PrismML. "With Bonsai 2 27B, we are closing the quality gap while keeping the same deployment advantages."
Model ships under Apache 2.0 license starting September 17, 2026
Developers and researchers can download Ternary Bonsai 2 27B for free under the Apache 2.0 license beginning September 17, 2026. PrismML emerged from a team of Caltech researchers with backing from Khosla Ventures, Cerberus, Google, and Samsung.