- Parameters
- 27B
- Quant
- Q2_g64 (1.71 bits/weight)
- Size
- 7.1 GB
- Family
- bonsai
ternary-bonsai-27b
Measured, not seateddenseresidentllama.cpp
Model card ↗Outcome as of 20 September 2026, from the seat registry. The write-up below was made when the model was tested.
Why it was tested
End-to-end ternary weights on a Qwen3.6-27B backbone with ~75% linear attention, claiming 262K context and 94.6% of FP16 intelligence retained. 11.7 GB resident at a 128k window - a third of K2-36B for the same context.
What came of it
Perfect tool use, collapsed epistemics. toolcall 100%, toolscale 100% (beating K2-36B), then worldview 42.5% and nullschema 27.8% with lure resistance at 12.5%. The clearest demonstration that a benchmark measuring only tool use would call a model lossless while half its judgement is gone. The upstream Q2_g64 variant lacks prism-ml's custom hybrid-attention kernels, so its 10.6 tok/s is a floor.
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| chainwalk | 0.0% | 0.0–65.8 | Passed 0/2 | llama.cppagentic-fast | 2026-09-14 |
| contract | 75.0% | 61.2–85.1 | Passed 36/48 | llama.cppextraction | 2026-09-14 |
| enrichment | 72.2% | 59.1–82.4 | Passed 39/54 | llama.cppextraction | 2026-09-14 |
| epistemics | 85.4% | 72.8–92.8 | Passed 41/48 | llama.cppstudy-nothink | 2026-09-14 |
| grounding | 68.8% | 60.2–76.3 | Passed 86/125 | llama.cppextraction | 2026-09-14 |
| nullschema | 27.8% | 18.8–39.0 | Passed 20/72 | llama.cppextraction | 2026-09-14 |
| screening | 93.3% | 84.1–97.4 | Passed 56/60 | llama.cppextraction | 2026-09-14 |
| synthesis | 66.7% | 54.1–77.3 | Passed 40/60 | llama.cppextraction | 2026-09-14 |
| tagger | 85.8% | 80.9–89.7 | Passed 206/240 | llama.cppextraction | 2026-09-14 |
| tagger_nullable | 90.0% | 85.6–93.2 | Passed 216/240 | llama.cppextraction | 2026-09-14 |
| toolcall | 100.0% | 94.0–100.0 | Passed 60/60 | llama.cppextraction | 2026-09-14 |
| tools | 100.0% | 79.6–100.0 | Passed 15/15 | llama.cppagentic-fast | 2026-09-14 |
| toolscale | 100.0% | 91.2–100.0 | Passed 40/40 | llama.cppagentic-fast | 2026-09-14 |
| worldview | 42.5% | 28.5–57.8 | Passed 17/40 | llama.cppextraction | 2026-09-14 |
Runs marked unenforced were measured through mlx_lm on a suite that requires constrained decoding. mlx_lm accepts the schema and ignores it, so those numbers do not measure what the suite is designed to measure. They are kept and labelled rather than deleted.
Every run log entry for this model (14)
| Finished | Suite | Runtime · profile | Score | Coverage | Grader · dataset · config |
|---|---|---|---|---|---|
| tagger_nullabletagger-nullable/1 | not recordedextraction | 90.0% | Coverage 240/240 | g 04e875431e26c4b5d db3addf23a2bc4cec 29960d736fce3ea5 | |
| taggertagger/2 | not recordedextraction | 85.8% | Coverage 240/240 | g 63e04d139a0e5ed3d 7ef587bf33e5788ac 773ed13c720ca034 | |
| enrichmentenrichment/3 | not recordedextraction | 72.2% | Coverage 54/54 | g b51222c2ca38cfdfd 75361864baa727abc 075725332e318d6b | |
| contractcontract/2 | not recordedextraction | 75.0% | Coverage 48/48 | g cc2d23c5bae45474d 1fe53d30cb5480f6c e5777a35ca1e5f10 | |
| chainwalkchainwalk/3 | not recordedagentic-fast | 0.0% | Coverage 2/2 | g f4bfcf6f040b0978d 129a9e74408c8919c 078f17fd21ba7d57 | |
| epistemicsepistemics/4 | not recordedstudy-nothink | 85.4% | Coverage 48/48 | g acf84972a0ad5f38d 4bce453fd376553cc 321e00a36a5faaeb | |
| groundinggrounding/1 | not recordedextraction | 68.8% | Coverage 125/125 | g 30b933dcc6a7e2d4d dcb396faf8b83b0ac 3b0afe72fb0c0121 | |
| screeningscreening/1 | not recordedextraction | 93.3% | Coverage 60/60 | g 0661b49a10be2d2ad 64c97ca2269c29afc c5097500538c3719 | |
| synthesissynthesis/2 | not recordedextraction | 66.7% | Coverage 60/60 | g 4156dd3c13361e8bd 267d1954e225c5b6c 6a368a22cfb7fb02 | |
| worldviewworldview/2 | not recordedextraction | 42.5% | Coverage 40/40 | g 9acfba1b704feb28d e97835de9e6fc227c 3a4970c1d0eed32d | |
| nullschemanullschema/1 | not recordedextraction | 27.8% | Coverage 72/72 | g d7448e0f5fc3f9afd 3b91280bb40372efc 004a863515b30c86 | |
| toolstools/1 | not recordedagentic-fast | 100.0% | Coverage 15/15 | g 2763675a5c3da564d 33a44812f0ad4c39c 7f71a1d06f005b6d | |
| toolscaletoolscale/1 | not recordedagentic-fast | 100.0% | Coverage 40/40 | g a2bc3751abfa28e0d b24abb108c409178c 13185468aa0f107f | |
| toolcalltoolcall/1 | not recordedextraction | 100.0% | Coverage 60/60 | g 804e8fc9457649d8d 10af88c170513155c 7dece85f852d6798 |
- Bonsai-27B (Q1_0)Rejected
26.9B · dense · Q1_0 (1.13 bpw) · 3.5 GB
Same base as the ternary build at a more aggressive operating point. Tested to find which capabilities survive 1-bit.
- Bonsai-8B (Q1_0)Rejected
8.0B · dense · Q1_0 (1.13 bpw) · 1.1 GB
Derived from Qwen3-8B, not the Qwen3.6-27B backbone the larger Bonsai models use. Tested as the small corner of a size-by-bit-width grid.
- Ternary-Bonsai-27B (PQ2_0)Measured, not seated
26.9B · dense · PQ2_0 (2.13 bpw, group-128) · 7.2 GB
prism-ml's group-128 ternary pack, built for their own llama.cpp fork's hybrid-attention kernels. The upstream Q2_g64 variant lacks them.