Sakura MicroQuality Kolibri-1 365E · file← Sakura MicroQuality Kolibri-1 365E
IQ3_XXS 27.45 GiB
The one we tested and kept, 29.5 GB. Available in our swap seat since 11 October. Webmp3 download; swap-seat bench entry.
- Size
- 29.5 GB
- Status
- benched; swap-seat bench entry sakura (live in fleetview)
- Scored runs
- 16
- Smallest Mac
- 48 GB estimate, 8K context
Results16 scored · 95% intervals
| Suite | Score | 95% interval | Passed | Runtime · profile | Recorded |
|---|---|---|---|---|---|
| chainwalksakura-iq3xxs-np4-b11533 | 50.0% | 9.5–90.5 | Passed 1/2 | llama.cppextraction | 2026-10-11 |
| contractsakura-iq3xxs-np4-b11533 | 75.0% | 61.2–85.1 | Passed 36/48 | llama.cppextraction | 2026-10-11 |
| enrichmentsakura-iq3xxs-np4-b11533 | 83.3% | 71.3–91.0 | Passed 45/54 | llama.cppextraction | 2026-10-11 |
| epistemicssakura-iq3xxs-np4-b11533 | 100.0% | 92.6–100.0 | Passed 48/48 | llama.cppextraction | 2026-10-11 |
| groundingsakura-iq3xxs-np4-b11533 | 79.2% | 71.3–85.4 | Passed 99/125 | llama.cppextraction | 2026-10-11 |
| needlesakura-iq3xxs-np2-128k-b11533 | 65.1% | 54.3–74.4 | Passed 54/83 | llama.cppextraction | 2026-10-11 |
| needlesakura-iq3xxs-np4-b1153350% coverage | 63.6% | 50.4–75.1 | Passed 35/55 | llama.cppextraction | 2026-10-11 |
| nullschemasakura-iq3xxs-np4-b11533 | 33.3% | 23.5–44.8 | Passed 24/72 | llama.cppextraction | 2026-10-11 |
| screeningsakura-iq3xxs-np4-b11533 | 95.0% | 86.3–98.3 | Passed 57/60 | llama.cppextraction | 2026-10-11 |
| synthesissakura-iq3xxs-np4-b11533 | 56.7% | 44.1–68.4 | Passed 34/60 | llama.cppextraction | 2026-10-11 |
| taggersakura-iq3xxs-np4-b11533 | 67.1% | 60.9–72.7 | Passed 161/240 | llama.cppextraction | 2026-10-11 |
| tagger_nullablesakura-iq3xxs-np4-b11533 | 85.0% | 79.9–89.0 | Passed 204/240 | llama.cppextraction | 2026-10-11 |
| toolcallsakura-iq3xxs-np4-b11533 | 100.0% | 94.0–100.0 | Passed 60/60 | llama.cppextraction | 2026-10-11 |
| toolssakura-iq3xxs-np4-b11533 | 86.7% | 62.1–96.3 | Passed 13/15 | llama.cppextraction | 2026-10-11 |
| toolscalesakura-iq3xxs-np4-b11533 | 100.0% | 91.2–100.0 | Passed 40/40 | llama.cppextraction | 2026-10-11 |
| worldviewsakura-iq3xxs-np4-b11533 | 65.0% | 49.5–77.9 | Passed 26/40 | llama.cppextraction | 2026-10-11 |
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 | Sakura-Kolibri-1-365E-IQ3_XXS-27.45GiB.gguf (29,472,035,488 B) |
| Format | IQ3_XXS mix, 3.17 bpw |
| Where from | downloaded 2026-10-11 |
| Status | benched; swap-seat bench entry sakura (live in fleetview) |
| Hugging Face | webmp3/Sakura-MicroQuality-Kolibri-1-365E-GGUF |
Where it came from
- Source:
webmp3/Sakura-MicroQuality-Kolibri-1-365E-GGUF, fileSakura-Kolibri-1-365E-IQ3_XXS-27.45GiB.gguf; revision not on record. - Size: 29,472,035,488 B (stat). sha256
cb53d1ca412ff83029b427a45656be725c2b15a00e8cd080c30e22780bceb788from the repo'sSHA256SUMS; the download script printed "sha256 OK". - Download:
pardl.sh, 6 ranges, started 2026-10-11 12:53Z, landed 13:23Z, sharing the line with the second Q8_0 pull. - Who asked: Alan, "grab the sakura version if it fits, always handy having models to test without fucking around".
- Quant: IQ3_XXS mix, 3.17 bpw; the GGUF's
file_typereads Q8_0 (inherited), and run manifests recordmodel_ftype: Q8_0for it (manifests) — do not infer the quant from that field.
How it was served
| label | dates (UTC) | binary and version | ctx / -np / KV / -fa / --jinja / --reasoning-format | harness profile | why this label exists |
|---|---|---|---|---|---|
sakura-iq3xxs-np4-b11533 | 2026-10-11 14:14-14:52Z (run log local 15:14:57-15:52:42) | llama.cpp-kolibri/build-master/bin/llama-server, 0.6.0-dev (build 11533, commit 769ced466) (manifest b11533-769ced466) | -c 131072 (32768 per slot) / 4 / f16 (no -ctk) / on / on / not set; --cache-ram 0 | extraction (thinking off), model-id kolibri-1-78b | the first Sakura gauntlet, four suites in flight, other five seats up (SUMMARY.md) |
sakura-iq3xxs-np2-128k-b11533 | 2026-10-11 16:03-17:03Z (local 17:03-18:03) | same binary | -c 262144 (131072 per slot) / 2 / f16 / on / on / not set; --cache-ram 0 | extraction | needle-only rerun after the 32K slots rejected 55 cases; new label so the 32K rows stay on record |
| — (no label; speed window) | 2026-10-11 13:43-13:48Z | same binary, seats down | llama-bench; batched-bench pp512/tg128 at 1/2/4 streams, f16 vs q4_0 KV | none | single-stream decode and cache A/B |
— (swap-seat use, sakura) | from 2026-10-11 ~19:50Z; reconfigured 2026-10-12 | same binary via fleetview /swap {"to":"sakura"} | PR #5: 4 x 32K f16; commit edb5acc: 2 x 64K; commit 607ac14: -ctk f16 -ctv f16 -c 262144 -np 2 --cache-ram 0 --alias sakura-kolibri | n/a (reasoning_effort: low default) | the live bench entry (fleetview.py) |
Seats held: swap seat on demand (an end-to-end test swapped sakura in (reported at 20:03Z) and the seat went back to ornith-9b); it was in the seat again on 2026-10-12 07:32Z. The gauntlet ran in the vacated seat with ornith-9b restored on exit.
Decisions
- Benched for the swap seat, not for a worker seat: the verdict was "a strong tool-and-screening occupant for the swap seat" and, of consolidating onto it, "Moving everything onto Sakura alone is a no".
- Registered as a play-seat bench entry on Alan's instruction "stick the sakura kolibri as an available model in the swap seat"; PR #5 merged 2026-10-11, fleetview restarted 19:16Z.
- Context raised twice on 2026-10-12 at Alan's prompting (see the family decision log).
What went wrong
- Needle at 32K slots (2026-10-11 14:15Z). 55 needle cases returned HTTP 400 ("request (33864 tokens) exceeds the available context size") because the assistant set
-c 131072 -np 4; the harness correctly scored them unmeasured. The published 63.6% (35/55) is a half-coverage number. - Stopped needle rerun. The 128K rerun was cut at 83/110 so the swap seat came back on schedule; the run-log row is 65.1% on 83 attempted (the 27 longest omitted).
- Locked suites broken in the repo. shapegen refused to start because 12 cases had changed without a version bump (
shapegen/2), and sitebuild consumes shapegen's data. Not a Sakura failure; no model can score either suite. - Resumed-run summary undercount. The Saluki run (same session) exposed that a resumed run prints only that session's rows; none of the Sakura files was resumed.
- Header/stamp traps (standing checks):
model_ftype: Q8_0in the manifests is inherited, not the quant. The run-logllama_buildb11533 (769ced466)is correct for this binary (commit 769ced466 is the rebased Hob-forge patch commit on upstream6184e92c5); upstream b11533 alone would not load the model.--reasoning-formatwas not set; profileextractionruns thinking off. - Slot config differs between the gauntlet and the live seat (4 x 32K scored; 2 x 128K served): no run-log label covers the live config except the needle-only
np2-128klabel.
What happened to it
Downloaded 2026-10-11 12:53-13:23Z -> speed window 13:43Z -> gauntlet 14:14-14:52Z and needle rerun 16:03-17:03Z -> bench entry live 19:16Z -> reconfigured 2026-10-12 -> on disk 2026-10-23. Not deleted; no RESTORE entry. Re-downloadable with models/kolibri-1-sakura-365e/download.sh.