Babak Hassibi's Caltech spinout says Bonsai 2 keeps 98.2% of full-precision performance, though its launch materials disagree on the test suite.
By [Ryan Merket](https://runtimewire.com/author/ryan-merket)
· Published
Primary source: [PR Newswire](https://www.prnewswire.com/news-releases/prismml-launches-bonsai-2-27b-its-most-capable-model-yet-302882228.html)
Why it matters #
PrismML is testing whether capable open-weight AI can move from rented cloud GPUs into private, local products. The footprint is meaningful; conflicting launch metrics make independent testing decisive.
PrismML launched Bonsai 2 27B on September 17th, giving developers an Apache 2.0-licensed version of Qwen3.8 27B whose language model fits into 5.95 GB. PrismML says its ternary weights preserve 98.2% of the full-precision model's benchmark performance while cutting the footprint by roughly nine times.
The release advances the central bet of co-founder and CEO Babak Hassibi, a Caltech professor who has spent his career working across information theory, signal processing, control and machine learning. Hassibi earned his doctorate in electrical engineering at Stanford, worked at Bell Labs and joined Caltech in 2001. PrismML has turned that mathematical background into a commercial argument: useful AI should be measured against the memory, energy and hardware required to deploy it.
Hassibi founded PrismML with Sahin Lale and Omead Pooladzandi, who serve as co-heads of research, and Reza Sadri, who serves as vice president of strategy. PrismML's team page also lists UC Berkeley professor Ion Stoica as an adviser. PrismML says it emerged from Caltech research and received backing or support from Khosla Ventures, Cerberus, Caltech and Google, with Samsung providing continuing support. Dealroom reported in April that PrismML had raised $16.25 million through a SAFE and seed financing from Khosla, Cerberus and Caltech; the round's valuation and ownership terms were not reported. Cerberus describes PrismML as a 2025-founded seed investment.
Three values instead of billions of precise ones
Bonsai 2 is derived from Qwen3.8 27B, an open-weight vision-language model with a native 262,144-token context window. PrismML represents most of its language weights using three possible values: -1, 0 and 1. Each group of 128 weights shares a higher-precision scale factor.
That representation brings the ideal language-model footprint to about 5.8 GB, compared with roughly 54 GB for the FP16 baseline. The most compact downloadable GGUF package is 5.95 GB, while a second packing designed to reduce unpacking work occupies 7.21 GB. The optional 8-bit vision component adds another 0.63 GB when image input is required.
An MLX companion is also available for Apple Silicon. The 5.9 GB figure is therefore a specific GGUF language-model configuration, rather than the size of every downloadable multimodal build.
PrismML also had to build the software path around the weights. The Bonsai 2 model card says the compact files require PrismML's fork of llama.cpp because the standard runtime does not support its custom ternary formats and Hadamard activation transform. PrismML has published that llama.cpp fork alongside a demo and integration repository.
That dependency narrows the meaning of drop-in local deployment. The weights are open and free, but developers must adopt PrismML's kernels and runtime build until the required formats reach widely used upstream inference software. PrismML says its backends cover CUDA, Metal and CPU execution.
The benchmark headline needs a footnote
PrismML's public launch materials agree on the 98.2% retention claim and disagree on several underlying figures.
The September 17th announcement reports an aggregate score of 83.9 across 20 benchmarks, compared with 85.4 for the full-precision model. It also describes Bonsai 2 as a 27.8-billion-parameter model capable of reaching 143 tokens per second on an Nvidia GeForce RTX 5090.
The model card published with the weights reports an average of 84.78 across 14 thinking-mode benchmarks, against 86.32 for FP16. It lists 27.36 billion total parameters and measured RTX 5090 generation throughput of 129.9 tokens per second for one packing and 120.5 for the smaller one. Those numbers may come from different evaluation or runtime configurations, but PrismML does not reconcile them in the published materials.
The model card supplies the more useful breakdown. Bonsai 2 scored 96.57 across four math tests, compared with 97.06 for FP16, and its coding average edged above the baseline, 89.42 to 89.07. Larger losses appeared in knowledge and reasoning, where Bonsai 2 scored 79.86 against 85.55, and vision, where it scored 66.19 against 71.36. On OCR Bench v2, the compressed model scored 56.88, below the FP16 model's 60.99 and a conventional 2-bit build's 61.70.
These are PrismML-run evaluations using PrismML's chosen tests, settings and aggregate formula. The release gives developers the weights needed to test whether the compression holds up in production workloads, where long contexts, tool errors, prompt processing and repeated agent loops matter more than a single blended score.
Hassibi's local-agent bet
PrismML released its first Bonsai 27B on July 14th, 2026. Based on Qwen3.6, that release included a 3.9 GB 1-bit variant aimed at phone-class deployment and a 5.9 GB ternary variant aimed at laptops. As of September 17th, Bonsai 2 27B is PrismML's current 27B flagship, using the newer Qwen3.8 base while retaining the 5.9 GB ternary footprint. PrismML says the earlier ternary model retained 95% of its full-precision benchmark average, compared with 98.2% for Bonsai 2.
Hassibi is aiming beyond private chatbots. PrismML identifies coding agents, computer-use systems, document analysis and long-running tool workflows as target applications. Local execution can keep files and intermediate agent state on the device while removing per-token API charges from repeated loops. A product can also route routine or sensitive steps to a local model and reserve cloud inference for requests that require greater capability.
The economics improve only when the compressed model remains dependable through those repeated steps. A small error rate can compound across a long agent run, and PrismML's own results show that capability losses vary materially by category. Vision and broad knowledge remain harder to compress than math and coding in PrismML's test suite.
Bonsai 2 gives developers a concrete artifact for testing Hassibi's thesis. It places a reasoning model with a 262K context window on a standard laptop and a single consumer GPU, with open weights and no usage fee. PrismML's next task is less mathematical: turning that footprint into applications whose users notice the privacy, speed and cost advantages without noticing the missing precision.