PrismML has released Ternary Bonsai 2 27B, a compressed version of Alibaba’s Qwen3.8 27B that the company says keeps 98.2 percent of the original model’s benchmark performance while shrinking its footprint by more than nine times, to 5.9GB. The claim rests entirely on PrismML’s own testing, and the company published no third-party evaluation alongside it.
The compression method is what PrismML calls ternary weighting: each parameter is stored as one of three values, -1, 0, or +1, combined with FP16 group-wise scaling. Each weight ends up carrying, on average, 1.76 bits of information, a ratio PrismML applies to the whole language model rather than to a handful of select layers. The result, according to the company, is a 27B-parameter model small enough to fit on a high-end laptop or gaming GPU, with a 262,000-token context window and support for both text and image input. Weights ship under the Apache 2.0 license, meaning developers can use and modify them without restriction.
PrismML frames the release as a sequel to its first Bonsai 27B model from two months earlier, which the company says retained about 95 percent of full-precision performance. Closing that gap to 98.2 percent is the headline improvement PrismML is selling, and on paper it narrows the tradeoff that has limited aggressively compressed models: giving up too much capability in exchange for a smaller footprint. On PrismML’s own benchmark suite, the compressed model scores 83.9 against Qwen3.8 27B’s 85.4, with the smallest gaps in instruction following and math and the largest in reasoning and vision tasks.
The company also reports throughput of 143 tokens per second on an Nvidia RTX 5090 and 46.8 tokens per second on Apple’s M5 Max chip, plus an energy figure of 0.714 milliwatt-hours per token on an RTX 4090, which PrismML says is 40 percent more efficient than an 8B model run at full precision. None of these figures come from an independent benchmark firm or a published third-party audit; they are PrismML’s measurements, on PrismML’s hardware selection.
The pitch is squarely about where AI runs, not just how capable it is. If a 27B-class model can genuinely hold onto most of its reasoning and coding ability at under 6GB, that changes what a laptop, phone, or edge device can do without calling out to a cloud API for every request. PrismML points to coding agents, computer-use workflows, and private document analysis as the target use cases, areas where sending every step to a remote model adds latency and cost.
That framing sits inside a broader shift in how compression gets pitched to developers: not as an emergency measure for underpowered hardware, but as a design choice for systems that mix local and cloud inference. PrismML, founded by a team of Caltech researchers with backing from Khosla Ventures, Cerberus, Google, and Samsung, is betting that low-bit models become the default deployment path rather than a fallback.
Teams evaluating Bonsai 2 27B for production agent loops should run their own benchmarks against Qwen3.8 27B before trusting the 98.2 percent retention figure, since PrismML has not published results from an outside lab.
PrismML detailed the release in its own launch post on prismml.com, published this month ahead of the September 18, 2026 run.
