Ternary-Bonsai-8B · file← Ternary-Bonsai-8B
PQ2_0
The file that loaded, fetched on 17 September and run on 13 of our 14 tests. Left running afterwards as a seat nobody used; deleted on 11 October. Prism-ml, prism fork.
- Size
- 2.2 GB
- Status
- Deleted 11 Oct 2026
- Scored runs
- 13
- Smallest Mac
- 8 GB estimate, 8K context
Results13 scored · 95% intervals
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| contractternary-bonsai-8b | 22.9% | 13.3–36.5 | Passed 11/48 | llama.cppextraction | 2026-09-17 |
| enrichmentternary-bonsai-8b | 70.4% | 57.2–80.9 | Passed 38/54 | llama.cppextraction | 2026-09-17 |
| epistemicsternary-bonsai-8b | 72.9% | 59.0–83.4 | Passed 35/48 | llama.cppstudy-nothink | 2026-09-17 |
| groundingternary-bonsai-8b | 51.2% | 42.5–59.8 | Passed 64/125 | llama.cppextraction | 2026-09-17 |
| nullschematernary-bonsai-8b | 11.1% | 5.7–20.4 | Passed 8/72 | llama.cppextraction | 2026-09-17 |
| screeningternary-bonsai-8b | 86.7% | 75.8–93.1 | Passed 52/60 | llama.cppextraction | 2026-09-17 |
| synthesisternary-bonsai-8b | 60.0% | 47.4–71.4 | Passed 36/60 | llama.cppextraction | 2026-09-17 |
| taggerternary-bonsai-8b | 43.8% | 37.6–50.1 | Passed 105/240 | llama.cppextraction | 2026-09-17 |
| tagger_nullableternary-bonsai-8b | 61.2% | 55.0–67.2 | Passed 147/240 | llama.cppextraction | 2026-09-17 |
| toolcallternary-bonsai-8b | 0.0% | -0.0–6.0 | Passed 0/60 | llama.cppextraction | 2026-09-17 |
| toolsternary-bonsai-8b | 93.3% | 70.2–98.8 | Passed 14/15 | llama.cppagentic-fast | 2026-09-17 |
| toolscaleternary-bonsai-8b | 100.0% | 91.2–100.0 | Passed 40/40 | llama.cppagentic-fast | 2026-09-17 |
| worldviewternary-bonsai-8b | 37.5% | 24.2–53.0 | Passed 15/40 | llama.cppextraction | 2026-09-17 |
Runs marked unenforced were measured on a suite that requires constrained decoding, on a server shown not to apply the schema: mlx_lm accepts it and ignores it, and a few other servers and models break it. Those numbers do not measure what the suite is designed to measure. They are kept and labelled rather than deleted.
The recordHow it was made · how it was served · what went wrong · what happened to it
Rebuilt from the lab's records by Claude, an AI model. The results above come straight from the automated runs. How these pages are made.
| File | Ternary-Bonsai-8B-PQ2_0.gguf, 2,182,184,672 B |
| Format | PQ2_0 GGUF |
| Where from | downloaded 2026-09-17 |
| Status | stray seat to 2026-09-19; deleted 2026-10-11 |
| Hugging Face | same repo |
Where it came from
- Source:
prism-ml/Ternary-Bonsai-8B-gguf, fileTernary-Bonsai-8B-PQ2_0.gguf. The repo lists F16, PQ2_0, Q2_0 and Q2_0_g64; the card page names only Q2_0, F16 and a Q2_0_G, so the PQ2_0 file "exists — the card just didn't list it" (assistant, listing). - Downloaded: 2026-09-17 07:46Z, complete by 07:48Z,
hf_hub_download(hf_transfer) intomodels/ternary-bonsai-8b-pq2/. Revision and sha256: not on record. - Runtime: the PrismML fork
llama.cpp-prism(b1-6ac5eb0, built 2026-09-14)./propsreportsmodel_ftype PQ2_0 - 2.13 bpw (group 128).
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
ternary-bonsai-8b | 2026-09-17 07:48Z launch; run-log 08:48:50 to 09:48:01 BST (07:48Z to 08:48Z); 13 suites (no chainwalk) | llama.cpp-prism/build/bin/llama-server, b1-6ac5eb0; run-log llama_build names stock b10893 (stale) | 32768 total, 4 slots (8192 per slot in the manifests) / 4 / q4_0 / on / on / not on the line (--cache-ram 1024 -ngl 999 --alias ternary-bonsai-8b) | extraction / agentic-fast / study-nothink | Fourth corner of the size x bit-width grid. Alan: "We never did manage to test the 8b bonsai". The same label had been used on 2026-09-15 for the void Q2_0 run. |
| (no label) speed probe | 2026-09-17 07:48Z | same | same | n/a | Two 300-token runs: 57.9 and 58.2 tok/s decode, prefill 156 then 54 tok/s. |
| (no label) stray seat | 2026-09-17 07:48Z to 2026-09-19 00:32Z | same | same | none | The server was left running after the gauntlet and later appeared in the fleet roster as bonsai-8b, "small ternary seat". |
Seats held: the server above, about 40 hours, as a four-slot 8K-per-slot utility seat. A 2026-09-17 13:57Z process listing still shows it, next to the production Ornith seat. It is not on record who added it to the roster; the assistant described it on 2026-09-19 as "sitting... since before this week", but it was started on 2026-09-17 for this gauntlet.
Decisions
- Why downloaded the first time (a different file): Ternary-Bonsai-8B Q2_0 (legacy Prism layout). This PQ2_0 file replaced it on 2026-09-17 once the load error named the layout.
- Purpose: to complete the 2x2 (27B/8B x ternary/1-bit) that tests whether a larger model loses more to compression; the result closed it: "The 2×2 is complete, and the finding holds decisively" (assistant). The recomputed numbers are below in Incidents.
- Alan on the seat (2026-09-19): the assistant told him this was the 8B Bonsai, its board record the weakest of anything resident, and k2-3.7b made it redundant. Alan: "Nothing uses it, was surprised to see it". The server was stopped and the roster line removed with the resurrection command preserved in a comment.
- Deleted 2026-10-11 with the Bonsai-1 family in the second sweep: the directory
ternary-bonsai-8b-pq2is listed as "retired seat variant (fleetview comment)" inRESTORE-deleted-20261011-tiersAB.sh, which also keeps the exact resurrection command.
What went wrong
- Label shared with a void run: the label
ternary-bonsai-8bfirst produced 1,032 rows of connection errors (void) on the Q2_0 file; those files were retired and the label reused for this valid run. The run-log rows are all from 2026-09-17; the retired files carry differentconfiguration_ids (results/_retired, 26 files). - The published 2x2 headline was computed from unfinished 8B files and from mean scores (published-data mismatch). The finding said the 27B loses 7.3 points to 1-bit and the 8B +1.3, "inside noise", over "12 paired suites". Replaying it from the result files: mean case score, the 12 suites that existed at 08:01Z (tagger_nullable not yet run on the 8B, chainwalk excluded), with the 8B ternary
taggertaken from its first 11 of 240 cases (86.4), gives exactly -7.32 (27B) and +1.26 (8B). On the finished files: mean case score over the 13 non-chainwalk suites -6.98 and +0.82; the board's pass-rate column -6.32 and -1.56 for the 8B over the 13 paired suites. By pass-rate the 8B loses 15.5 points ontagger(43.8 to 28.3) and 19.5 ontagger_nullable(61.2 to 41.7). The conclusion that the larger model loses more survives; "loses nothing" does not on the board's own column. - toolcall 0/60: all 60 result rows have outcome
answered_directly(result file read 2026-10-23). It is inherited from the Qwen3-8B base (its midrange run is also 0/60) while toolscale is 100: the model formats a call perfectly and does not decide to make one (Tool perception is mostly all-or-nothing and tracks the checkpoint, not size or vendor claims). The assistant's suspicion that it was "a harness-era artefact" is not supported: the base shows the same. - Stale build stamp / missing flags: as for the other prism-fork runs;
--jinjaon, no--reasoning-format.
What happened to it
Downloaded 2026-09-17 → benched 2026-09-17 → stray seat to 2026-09-19 → retired 2026-09-19 → deleted 2026-10-11 (RESTORE-deleted-20261011-tiersAB.sh, which also calls the sibling directory ternary-bonsai-8b "benched: 14 rows; RETIRED 2026-09-19 seat": the 14 rows belong to this file, not the Q2_0 directory). Re-downloadable from prism-ml/Ternary-Bonsai-8B-gguf.