PrismML's Ternary Bonsai 2 27B compresses a 27-billion-parameter model to under 6GB while retaining 98.2% of baseline performance.
The AMW Read
Second-generation ternary quantization from a known Apple-acquisition-rumored player meaningfully improves benchmark retention (98.2% vs ~95%) via an explicit open-weight release strategy.
PrismML's Ternary Bonsai 2 27B compresses a 27-billion-parameter model to under 6GB while retaining 98.2% of baseline performance.
PrismML, a Caltech spinoff previously linked to Apple acquisition talks after shrinking a 54GB Qwen model to under 4GB for on-device use on the iPhone 17 Pro in July, released Ternary Bonsai 2 27B on September 17. Built on Alibaba's open-source Qwen2.5 27B, the model replaces standard 16-bit weights with a ternary scheme using only three values (-1, 0, +1) combined with FP16 group-level scaling, shrinking the 53.8GB FP16 original to 5.93GB — about 1.76 effective bits per parameter. Across 20 benchmarks it retained 98.2% of the original model's performance, up from roughly 95% for the first Bonsai 27B two months earlier, though it still lags on multi-step agentic work: 52.8 versus 69.7 on Terminal-Bench 2.1, and 60.8 versus 80.6 on SWE-bench Verified.
The release sits in the on-device AI compression race, where the binding constraint is whether a capable model fits in a phone's or laptop's memory rather than raw capability. PrismML's ternary weighting pushes compression further than typical INT4/INT8 quantization while narrowing the accuracy gap that has usually made aggressive compression unusable outside toy benchmarks. Multimodal input, a 262K-token context window, and native CUDA and Apple MLX execution paths target both edge GPU and consumer-device deployment — extending, not just repeating, the iPhone demo that first drew Apple's interest.
A 27B-class model under 6GB running at 142.5 tokens/second on an RTX 5090 and 46.8 on an Apple M5 Max makes locally run coding assistants and document or computer-control agents workable without server round-trips, provided the task avoids long agentic chains, where the compression penalty persists. The Apache 2.0 release on Hugging Face, paired with standing Apple acquisition speculation, points to a company monetizing licensing interest rather than API revenue; per the AI Market Watch index, the Caltech spinoff has raised $16.25M to date (coverage, not a census) — a modest base for a startup now positioned as compression IP a hyperscaler might want to own outright.