PrismML ships Ternary Bonsai 2 27B, 9x smaller at 98.2% retention
PrismML released Ternary Bonsai 2 27B, the second model in its Bonsai series and a follow-up to the original Bonsai 27B it shipped two months earlier. The new model is built on Qwen3.8 27B and uses ternary weights, restricted to the values -1, 0 and +1, combined with FP16 group-wise scaling. That works out to 1.76 effective bits per weight and a total model footprint of 5.9GB. It supports a 262K-token context window, takes both text and image input, and ships under the Apache 2.0 license.
Across a benchmark suite covering reasoning, math, coding, instruction following, vision and agentic tool use, Ternary Bonsai 2 27B scores 83.9 in thinking mode, which PrismML says is 98.2% of the aggregate score the full-precision Qwen3.8 27B achieves. Against that same full-precision counterpart, the compressed model is described as more than 9x smaller. PrismML says the retained capability is concentrated in exactly the areas most sensitive to degradation: coding agents, tool-use systems, multimodal workflows and long-horizon tasks, where small errors compound over many steps.
Compared with the original Ternary Bonsai 27B released two months prior, which retained about 95% of full-precision capability, PrismML says Bonsai 2 27B closes that gap to over 98%, a level the company calls practically lossless. On hardware, PrismML reports throughput of up to 143 tokens per second on an NVIDIA GeForce RTX 5090 and 46.8 tokens per second on an Apple M5 Max, and energy use of 0.714 mWh per token on an RTX 4090, which it says makes the model 40% more energy-efficient than an 8B model running at full precision.
The model runs on NVIDIA GPUs via CUDA and on Apple devices (Mac, iPhone, iPad) via MLX, through custom low-bit kernels, with weights available now under Apache 2.0 and further benchmarking detail in an accompanying whitepaper. PrismML says it emerged from a team of Caltech researchers and was founded with support from Khosla Ventures, Cerberus and Google, with continuing support from Samsung.
Key facts
- Ternary Bonsai 2 27B is built on Qwen3.8 27B, uses ternary {-1, 0, +1} weights with FP16 group-wise scaling for 1.76 effective bits per weight, and has a 5.9GB total footprint with a 262K-token context window.
- It scores 83.9 on a reasoning, math, coding, vision and agentic-tool-use benchmark suite in thinking mode, retaining 98.2% of full-precision Qwen3.8 27B's aggregate performance while being more than 9x smaller.
- Versus the original Ternary Bonsai 27B from two months earlier, PrismML says it closed the capability-retention gap against full precision from 95% to over 98%.
- Reported throughput reaches up to 143 tokens per second on an NVIDIA GeForce RTX 5090 and 46.8 tokens per second on an Apple M5 Max; on an RTX 4090 it uses 0.714 mWh per token, 40% more energy-efficient than an 8B model running at full precision.
- Weights are released under Apache 2.0, running on NVIDIA GPUs via CUDA and Apple devices via MLX; PrismML says it emerged from a team of Caltech researchers with backing from Khosla Ventures, Cerberus, Google and Samsung.
Why it matters
PrismML frames this release around closing a gap: its first Bonsai 27B model retained about 95% of full-precision capability after compression, and this second version pushes that to over 98%, which the company calls practically lossless. At that retention level, PrismML argues, extreme compression stops being a tradeoff and becomes a deployment unlock: a model can be shrunk by more than 9x while giving up almost nothing on a broad benchmark suite. PrismML frames the implications beyond a single local model, saying low-bit compression can change memory and energy economics across devices, workstations and datacenters, not just phones and laptops.
Who it affects
The release targets developers and teams building AI systems with tight memory, latency or power budgets: coding assistants that need fast edit-debug loops, multimodal agents processing screenshots and documents, and privacy-sensitive local workflows that want to avoid constant cloud calls. It also affects hardware owners specifically named in PrismML's numbers, since the reported throughput and energy figures are tied to particular GPUs and Apple silicon (RTX 5090, RTX 4090, M5 Max) rather than being generic claims.
How to use it
Model weights for Ternary Bonsai 2 27B are available now under the Apache 2.0 license. It runs on NVIDIA GPUs via CUDA and on Apple devices, Mac, iPhone and iPad, via MLX, using PrismML's custom low-bit kernels. PrismML points to a whitepaper for full technical detail on the compression, evaluation and benchmarking process, and says it will work with teams to tailor Bonsai models to specific applications, including post-training on domain data and inference optimization for target hardware.
How solid is it
Every performance figure here, the 98.2% retention, the 83.9 benchmark score, the throughput and energy numbers, comes from PrismML's own announcement and its own whitepaper; the source text gives no sign of independent or third-party verification of these results. The per-benchmark breakdown behind the aggregate 83.9 score is referenced as sitting in a figure and the whitepaper rather than being laid out in the announcement itself, so the aggregate number cannot be checked against its components from this text alone.
Risks and caveats
The announcement does not name the "other low-bit alternatives" it says Bonsai 2 27B outperforms on intelligence density, so that comparison cannot be checked against specific competitors. No pricing or API-access terms are given beyond the fact that weights are open under Apache 2.0, and no absolute release date is stated for either Bonsai release, only that the first came two months before this one.
“This is a significant improvement that makes the current release practically "lossless". It further cements the notion that low-bit models can be the best way to deploy AI.”
— PrismML, in the Bonsai 2 27B announcement