PrismML Debuts Bonsai 2 27B, a Lean Open‑Source LLM Built for Efficiency

U.S. startup PrismML announced a new large language model, Bonsai 2 27B, on September 17. The roughly 27‑billion‑parameter model is derived from Alibaba’s Qwen 3.8 27B architecture but has been heavily tuned for efficient inference, making it easier to deploy outside hyperscale data centers.

PrismML describes Bonsai 2 27B as a “ternary” model, using a three‑level numerical representation that compresses weights beyond common FP16 or FP8 formats. The aim is to cut power consumption and memory usage while preserving most of the model’s accuracy. That could make Bonsai 2 27B attractive not only for GPU clusters, but also for on‑premise servers and high‑end workstations where organizations want tight control over their AI stack.

The model is released under the permissive Apache 2.0 license, allowing researchers and companies to modify and commercialize it without royalties. Performance sits in the mid‑range of today’s open models—strong enough for many enterprise workloads, but not meant to compete head‑on with the largest frontier systems. Instead, PrismML is positioning Bonsai 2 27B as a practical, self‑hostable option for teams that care about cost, latency, and data sovereignty as much as raw benchmark scores.