PrismML

Overview

A compression lab, described in one source as a Caltech spin-out, that publishes ternary-weight versions of other labs' open-weight models under permissive licences. Its company line is "Concentrating intelligence" (source).

It is the first entity on this wiki whose product is a transformation of a model rather than a model. Sakana AI sells a model it did not train — the Fugu line routes to other labs' models behind one endpoint — but it trains the orchestrator. PrismML trains neither the base model nor a router: its stated contribution is the ternary representation and the quantization-aware training that makes it survivable, applied to somebody else's weights (source). What it ships is a 9× smaller copy of Qwen 3.8 27B.

Nothing read gives its founding date, headcount, funding or any prior release except the first-generation Bonsai line, which is listed on Together AI as prism-ml-ternary-bonsai-27b. The Bonsai 2 name implies a predecessor and this wiki has no page for it.

Key People

None named. No source read gives a founder, CEO or research lead (source). Recorded as unknown rather than left implied.

Models & Products

• Ternary Bonsai 2 27B — released 2026-09-17, Apache 2.0, 5.93 GB against 53.80 GB in FP16, 1.76 effective bits per weight, 262,144-token context, text and image input. Claimed 98.2% of the parent's 20-benchmark aggregate
• Ternary Bonsai 27B (first generation) — no release date, specification or benchmark in anything read; known only from a Together AI listing and two Hugging Face repositories. No page, because there is nothing to write one from

Recent Activity

• 2026-09-17: Released Ternary Bonsai 2 27B, and the interesting number is the one under the headline. A ternary {−1, 0, +1} quantization of Qwen 3.8 27B with FP16 group-wise scaling, 1.76 effective bits per weight, 5.93 GB total against 53.80 GB for the FP16 parent, Apache 2.0, GGUF and MLX on Hugging Face, claimed to run on a 16 GB laptop or a single 24 GB GPU (3 passes). The headline is 98.2% retention — an aggregate of 83.9 against the parent's 85.4 across 20 benchmarks spanning reasoning, math, coding, instruction following, vision and agentic tool use. On Terminal-Bench 2.1, which sits outside that suite, it scores 52.8 against the parent's 69.7 — about 75% (1 pass), and the same pass puts SWE-bench Verified in the same band without giving figures. Why it matters for this page: the two numbers are both PrismML's own and they are not in tension — they measure different things — but only one of them is in the headline, and the one that is omitted is the one that describes long-horizon agent work. Method claim against conventional low-bit builds: 1.23× smaller than IQ2_XXS GGUF and 8.7 points higher, which is the entire argument for quantization-aware ternary training over post-training quantization (1 pass). No first-party read — prismml.com answers EGRESS_BLOCKED, new to this repo's blocked list, as does huggingface.co → Ternary Bonsai 2 27B, Alibaba / Qwen AI Lab, Open-Weights Policy Fight (source) (MarkTechPost)

Strategic Position

Its position depends entirely on somebody else shipping open weights, and it says so by existing. The reason Qwen 3.8 27B exists at all is that Qwen 3.8 Max's smallest published quantisation is 397 GB; the reason PrismML exists is that 53.80 GB is still more than a laptop has. Each step down the size ladder is taken by a different party, and only the first one trained anything.

That makes Apache 2.0 on the parent a precondition, not a preference. PrismML can publish a derivative of Qwen 3.8 27B under Apache 2.0 because Alibaba published the parent that way. The same transformation applied to a closed model has no legal path, which is the practical shape of Open-Weights Policy Fight this wiki has otherwise only recorded in the abstract.

What is not established: no price, no hosted endpoint for Bonsai 2, no throughput or energy figure despite both being claimed, and no third-party evaluation of any PrismML artefact of any kind (source).

Related

• Alibaba / Qwen AI Lab — trains the parent model; the whole of PrismML's published output derives from one of its checkpoints
• Open-Weights Policy Fight — the licence terms that make a third-party derivative publishable
• Test-Time Compute (Inference-Time Compute Scaling) — the other axis on which this wiki tracks what a small local model actually costs to run
• Liquid AI — the nearest comparison by intent: on-device open weights, reached by training small rather than by compressing large