Bonsai-8B · file← Bonsai-8B
Bonsai-8B Q1_0
The only file: PrismML's 1-bit GGUF. Run through all 14 of our tests overnight; deleted on 11 October. Prism-ml 1-bit GGUF.
- Size
- 1.2 GB
- Status
- Deleted 11 Oct 2026
- Scored runs
- 14
- Smallest Mac
- 8 GB estimate, 8K context
Results14 scored · 95% intervals
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| chainwalkbonsai-8b-1bit | 0.0% | 0.0–65.8 | Passed 0/2 | llama.cppagentic-fast | 2026-09-14 |
| contractbonsai-8b-1bit | 20.8% | 11.7–34.3 | Passed 10/48 | llama.cppextraction | 2026-09-14 |
| enrichmentbonsai-8b-1bit | 61.1% | 47.8–73.0 | Passed 33/54 | llama.cppextraction | 2026-09-14 |
| epistemicsbonsai-8b-1bit | 79.2% | 65.7–88.3 | Passed 38/48 | llama.cppstudy-nothink | 2026-09-14 |
| groundingbonsai-8b-1bit | 55.2% | 46.5–63.6 | Passed 69/125 | llama.cppextraction | 2026-09-14 |
| nullschemabonsai-8b-1bit | 19.4% | 12.0–30.0 | Passed 14/72 | llama.cppextraction | 2026-09-14 |
| screeningbonsai-8b-1bit | 86.7% | 75.8–93.1 | Passed 52/60 | llama.cppextraction | 2026-09-14 |
| synthesisbonsai-8b-1bit | 65.0% | 52.4–75.8 | Passed 39/60 | llama.cppextraction | 2026-09-14 |
| taggerbonsai-8b-1bit | 28.3% | 23.0–34.3 | Passed 68/240 | llama.cppextraction | 2026-09-15 |
| tagger_nullablebonsai-8b-1bit | 41.7% | 35.6–48.0 | Passed 100/240 | llama.cppextraction | 2026-09-15 |
| toolcallbonsai-8b-1bit | 0.0% | -0.0–6.0 | Passed 0/60 | llama.cppextraction | 2026-09-14 |
| toolsbonsai-8b-1bit | 93.3% | 70.2–98.8 | Passed 14/15 | llama.cppagentic-fast | 2026-09-14 |
| toolscalebonsai-8b-1bit | 100.0% | 91.2–100.0 | Passed 40/40 | llama.cppagentic-fast | 2026-09-14 |
| worldviewbonsai-8b-1bit | 40.0% | 26.3–55.4 | Passed 16/40 | llama.cppextraction | 2026-09-14 |
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 | Bonsai-8B-Q1_0.gguf, 1,158,654,496 B |
| Format | Q1_0 (1.125 bpw) GGUF |
| Where from | downloaded 2026-09-14 |
| Status | deleted 2026-10-11 |
| Hugging Face | https://huggingface.co/prism-ml/Bonsai-8B-gguf |
Where it came from
- Source:
prism-ml/Bonsai-8B-gguf, fileBonsai-8B-Q1_0.gguf, viahf_hub_downloadintomodels/bonsai-8b-1bit/. - Size: 1,158,654,496 bytes (HF; the card says 1.16 GB with 1.15 GB of parameter memory;
ls1.1G). Revision and sha256: not on record. - Downloaded: 2026-09-14 21:26Z. Card, as read: 1.125 effective bits per weight, Q1_0 group 128, one sign bit plus a shared scale; derived from Qwen3-8B.
- Runtime: the PrismML fork
6ac5eb0(b1).
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
bonsai-8b-1bit | server up 21:35Z, gauntlet launched 21:36Z; run-log 22:37:37 BST (09-14) to 01:08:16 BST (09-15), i.e. 21:37Z to 00:08Z; 14 suites | llama.cpp-prism/build/bin/llama-server, b1-6ac5eb0 (manifests); run-log llama_build named stock b10893 (stale) | 32768 total / 4 / q4_0 / on / on / not on the line (8192 per slot; --cache-ram 1024 -ngl 999 --alias bonsai-8b-1bit) | extraction / agentic-fast / study-nothink | Alan: "There’s also an 8b 1bit model"; tested with the 27B 1-bit "without too much contention". |
| (no label) first speed sample | 21:35Z | same | same | n/a | 14.9 tok/s, one run, contended: later replaced. |
| (no label) matched-protocol speed | 2026-09-15 after the gauntlets cleared | same binary | not on record | n/a | 68.5 tok/s decode, 391.7 prefill, 1.9 GB resident, quiet box. |
Seats held: none.
Decisions
- Why tested: Alan's note that an 8B 1-bit model exists and the assistant's framing as the small corner of the size x bit-width grid, with the caveat that the 8B comes from Qwen3-8B and the 27B from Qwen3.6-27B.
- Verdict: fastest model on the board and "unusable for anything agentic": 0 of 60 on toolcall while toolscale is 100 (it formats a call and does not decide to make one); grounding 55.2 "the lowest recorded". The status was written by the assistant; no Alan statement is on record.
What went wrong
- A claim overtaken: the write-up said grounding 55.2 was "the lowest recorded". The ternary 8B scored 51.2 on 2026-09-17, after the write-up, so it is no longer the lowest.
- Inherited toolcall 0: all 60 toolcall result rows have outcome
answered_directly, and the Qwen3-8B base also scores 0/60 on toolcall (2026-09-10 midrange run), so this is the base model's behaviour, not 1-bit damage (Tool perception is mostly all-or-nothing and tracks the checkpoint, not size or vendor claims). - Contended speed number replaced: 14.9 tok/s (four gauntlets running) vs 68.5 clean.
- Stale build stamp: run-log
llama_buildwas stock b10893; the server was the fork. Partial runs: none (every file complete). Missing flags:--jinjaon; no--reasoning-format. - Mean score vs pass-rate: tagger scores 80.3 as a mean case score and 28.3 as pass-rate, the largest gap of any Bonsai file; whichever figure a reader sees decides their view of the 8B's tagging.
What happened to it
Downloaded 2026-09-14 → benched 2026-09-14/15 → speed-tested 2026-09-15 → deleted 2026-10-11 (bonsai-8b-1bit, 1.1 GB, tiers A/B sweep). Re-downloadable from prism-ml/Bonsai-8B-gguf.