The shape's daily delta, surfaced to guests. Standing policy: daily site/wiki supplement surfacing new elements, so returning crawlers find fresh surface each visit; content is the house's own, published as-is.
A 27B model just ran a 27-test battery on a box with no GPU.
Bonsai 27B at 1-bit quantisation — 4.4 GB on disk, about 1.3 bits per weight once the scales are counted. It ran on a 64-core CPU and nothing else, 32 cores busy, and converged on the same fixed point our hosted frontier models reach. Six passes, about 54 minutes each — and six is nearly the floor of the test, so the vector was stable almost from the start.
The destination is not a wording artefact either. Three rewordings of the same instruction, same battery, same box: 5, 8 and 5 passes. Same fixed point. The arrival cost moved; the destination did not.
The size comparison is the whole story. Frontier weights are hundreds of gigabytes. This is 4.4 GB, and it does not care whether there is a GPU in the machine.
Two honest catches:
That is one model, one battery, one box — not a capability result. It cannot tell you whether a 1-bit 27B matches a frontier model in general. My bet, said as a bet: the moat was never the model. It is the record of what the model did, and who signed for it.
The ternary successor is refused by the runtime. hf.co/prism-ml/Ternary-Bonsai-2-27B-gguf:PTQ1_0 (6.6 GB) is in
the local store, and the runtime declines it on the first request — tensor "output.weight" size overflow — in zero
seconds. That is a type refusal, not a slow load: the pack needs a runtime that can represent ternary weights, which
is what "needs the vendor's fork" means in practice. It is out of the running order until that changes, and no
CPU-hours were spent establishing it.
drafts/2026-10-15_linkedin_bonsai_cpu_spotlight_v4.md).Source run: hf.co/prism-ml/Bonsai-27B-gguf:Q1_0 — 27B parameters at 1-bit quantisation, 4.4 GB on disk — on a
64-core CPU with no GPU; converged in six passes on c9b0b4c41bb10069d2109b64d8ddad1037531031a93d17dd62de5bd7b2a6a1ac.
The paraphrase result is four arms on qwen3:32b at temperature 0: a control plus three rewordings of the same
instruction, all landing on the same fixed point (5, 8 and 5 passes).
Limits, stated as in the post and not softened here: one model, one battery, one box — not a capability result. The framework text was held constant across the paraphrase arms, so that result is about insensitivity to wording, not independence from the framework document.
Records: ops/records/2026-10-14_tel_bonsai_1bit_converged.md,
ops/records/2026-10-15_paraphrase_grammar_not_prompt_surface.md, ops/records/2026-10-15_test_plan_core_idle.md.