Ornith-1.5-9B · file← Ornith-1.5-9B

Ornith-1.5-9B Q4_0

The one we kept, and the only file we held. A community 4-bit build; it ran our agents for a day and a half. Bartowski download.

Size: 5.5 GB
Status: On disk
Scored runs: 13
Smallest Mac: 16 GB estimate, 8K context

Results13 scored · 95% intervals

Suite	Score	95% interval	Passed	Runtime · profile	Recorded
contractornith-1.5-9b	68.8%	54.7–80.1	Passed 33/48	llama.cppextraction	2026-09-12
epistemicsornith-1.5-9b	100.0%	92.6–100.0	Passed 48/48	llama.cppextraction	2026-09-12
groundingornith-1.5-9b	81.6%	73.9–87.4	Passed 102/125	llama.cppextraction	2026-09-12
groundingornith-9b-kvq4	80.0%	72.1–86.1	Passed 100/125	llama.cppextraction	2026-09-12
needleornith-9b-kv-f16	58.2%	48.8–67.0	Passed 64/110	llama.cppextraction	2026-09-12
needleornith-9b-kv-q40	55.5%	46.1–64.4	Passed 61/110	llama.cppextraction	2026-09-13
nullschemaornith-1.5-9b	52.8%	41.4–63.9	Passed 38/72	llama.cppextraction	2026-09-12
nullschemaornith-9b-kvq4	56.9%	45.4–67.7	Passed 41/72	llama.cppextraction	2026-09-12
synthesisornith-1.5-9b	80.0%	68.2–88.2	Passed 48/60	llama.cppextraction	2026-09-12
toolcallornith-1.5-9b	100.0%	94.0–100.0	Passed 60/60	llama.cppextraction	2026-09-12
toolcallornith-9b-kvq4	100.0%	94.0–100.0	Passed 60/60	llama.cppextraction	2026-09-12
toolscaleornith-1.5-9b	95.0%	83.5–98.6	Passed 38/40	llama.cppextraction	2026-09-12
worldviewornith-1.5-9b	62.5%	47.0–75.8	Passed 25/40	llama.cppextraction	2026-09-12

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	`Ornith-1.5-9B-Q4_0.gguf` (+ `mmproj-Ornith-1.5-9B-f16.gguf`, 918,165,728 B)
Format	Q4_0, 5,482,832,768 B
Where from	downloaded 2026-09-12 ~09:21Z from `bartowski/Ornith-1.5-9B-GGUF`
Status	on disk; occupant of the swap seat since 2026-09-28 (last transcript record: 2026-10-11)
Hugging Face	https://huggingface.co/bartowski/Ornith-1.5-9B-GGUF

Where it came from

Repo and file: bartowski/Ornith-1.5-9B-GGUF, Ornith-1.5-9B-Q4_0.gguf and mmproj-Ornith-1.5-9B-f16.gguf. The official ornith-ai/Ornith-1.5-9B-GGUF was listed in the same check and not used.
Command (2026-09-12 ~09:21Z; hfpull.sh is a scratchpad range-request downloader, 6 streams for the weights, 2 for the projector):

mkdir -p models/ornith-1.5-9b
hfpull.sh bartowski/Ornith-1.5-9B-GGUF Ornith-1.5-9B-Q4_0.gguf models/ornith-1.5-9b/Ornith-1.5-9B-Q4_0.gguf 6
hfpull.sh bartowski/Ornith-1.5-9B-GGUF mmproj-Ornith-1.5-9B-f16.gguf models/ornith-1.5-9b/mmproj-Ornith-1.5-9B-f16.gguf 2

Download finished clean between 09:22Z and 09:28Z. File mtimes 12 Sep 10:24-10:25 local (09:24-09:25Z).

Size: 5,482,832,768 B (stat, 2026-10-23); mmproj 918,165,728 B. The 5.48 GB figure was quoted from the HF listing before download. sha256 and HF revision: not on record (would have been in the download log, which hfpull.sh did not keep). Format check: GGUF general.file_type = 2 (Q4_0).
Who / why: Alan, "Yes do the 9b"; see the family record.
Not made here: no quantisation was run on this model; the file is the vendor-community Q4_0 as downloaded.

How it was served

label	dates (UTC)	binary and version	ctx / -np / KV / -fa / `--jinja` / `--reasoning-format`	harness profile	why this label exists
`ornith-1.5-9b` (toolcall, toolscale, nullschema, grounding, worldview)	2026-09-12 09:31-09:39Z	`.local/bin/llama-server`, b10893-c32d1dabe (run-log `llama.cpp b10893 (c32d1dabe)`, runtime_source `props`)	-c 32768, 4 slots (manifest n_slots 4, so 8,192 per slot), f16 KV, `-fa on`, `--jinja`, no reasoning-format flag, `--mmproj`, `-ngl 99` (manifests)	extraction	the retest of Ornith as a Hermes driver (family record)
`ornith-1.5-9b` (synthesis, contract, epistemics)	2026-09-12 16:31-16:37Z	same binary	-c 131072, -np 1, f16 KV, `-fa on`, `--jinja`, `--cache-ram 16384`, mmproj on (`swap-to-ornith9b.sh` as rewritten 11:59Z; manifests n_ctx 131072, n_slots 1)	extraction	cover the retrieve-then-synthesise half of the workload the first batch had not measured. Ran against the seat Alan was using for Hermes
`ornith-9b-kvq4`	2026-09-12 12:00-12:06Z	same	-c 262144, -np 1, `-ctk q4_0 -ctv q4_0`, `-fa on`, `--cache-ram 16384`, alias `ornith-9b-kvq4`	extraction	does q4_0 KV cost quality at the full native window? Shallow suites only (nullschema, grounding, toolcall)
`ornith-9b-kv-f16`	2026-09-12 19:38-21:33Z (resumed; 12 cases carried over from aborted launches)	same	-c 131072, -np 1, f16 KV (live driver; manifest `cache_type_k` is null because /props did not report it)	extraction	needle/1 arm 1: f16 KV baseline
`ornith-9b-kv-q40`	2026-09-12 ~21:40-23:28Z	same	-c 49152, -np 1, `-ctk q4_0 -ctv q4_0`, `-fa on`, `--cache-ram 8192`, alias `ornith-9b-kvq4` (the `--model-id`) while the label is `ornith-9b-kv-q40`	extraction	needle/1 arm 2: q4_0 KV, paired with arm 1 on the same 110 cases

Command for the first label:

.local/bin/llama-server -m models/ornith-1.5-9b/Ornith-1.5-9B-Q4_0.gguf --mmproj models/ornith-1.5-9b/mmproj-Ornith-1.5-9B-f16.gguf \
  -c 32768 -ngl 99 --jinja -fa on --alias ornith-1.5-9b

and for the harness, per suite: harness/runner.py --suite <s> --model <server> --model-id ornith-1.5-9b --label ornith-1.5-9b --profile extraction.

Seats held (outside the run log):

Hermes driver and all 11 auxiliary roles, 2026-09-12 10:08Z to 2026-09-13 ~20:15Z (stopped when Alan asked for just Qwen3.8). First at -c 65536 -np 1 -fa on; restarted 11:59Z at -c 131072 --cache-ram 16384; Hermes pointed at it with default: ornith-1.5-9b.
Concurrency for Spumco's cron tests, 2026-09-13: 4 slots x 16,384 (19:19Z, 8.3 GB), then -c 65536 with -np 4 (19:55Z, 8.4 GB).
Alias only, 2026-09-13 20:47Z onward: the seat answered to ornith-1.5-9b but served Qwen3.8-27B.
swap-seat occupant, 2026-09-28 15:44Z to the last record (2026-10-11): hand-started on llama-server -m Ornith-1.5-9B-Q4_0.gguf --jinja -ctk q4_0 -ctv q4_0 -fa on... (command line truncated in the record), parent = fleetview, launched by swap-to-ornith9b.sh per the 2026-10-09 omp summary. Vacated on 2026-10-11 for the Sakura and Saluki gauntlets and restored each time (monitor event 'restoring the seat -> ornith-9b' at 14:52Z).
Registered in fleetview BENCH["ornith-9b"]: -c 131072 -np 1 --alias ornith-9b, label "Ornith 1.5 9B dense Q4_0".

Decisions

Downloaded to answer whether the dense member of the family could drive Hermes where the MoE's v2 scores looked poor.
Kept and seated as the Hermes driver because of tool reach (100% toolcall) and size; the cost known going in was nullschema 52.8% and worldview 62.5%, with the instruction not to ask it to declare absences or cite sources.
Driver seat stays on f16 KV. From the needle result: the driver feeds FreshLoops and the graph, where admission is citation-grounded, so the 14.7 GB q4_0 would buy is paid for in citation accuracy. kv-q40-driver.sh refuses without --accept-citation-cost (script header; finding kv-quantisation-degrades-citation-not-recall).
Displaced as Hermes driver on 2026-09-13 ~20:12Z by Qwen3.8-27B (see the family decision log). The 35B-A3B was a separate seat and is not what displaced it.
Not retired, not deleted: the file survived the 2026-10-11 sweeps; it was then the swap seat's occupant, and it is fleetview's ornith-9b bench entry.

What went wrong

Partial worldview quoted as a result (2026-09-12 09:38Z). The assistant reported worldview 82.1% from a 28-case partial. The full run finished at 62.5% (25/40) and the assistant said so at 09:40Z. The run log holds only the full run.
Needle suite launches that failed or were killed (2026-09-12 19:19-19:46Z). Four attempts: (a) refused by the lock, because needle/1 was not frozen; (b) KeyError: 'id' at case zero (the suite used case_id); (c) a foreground sleep 120 hit the 2-minute tool timeout and killed the process group, cancelling requests mid-prefill; (d) setsid does not exist on macOS. The f16 manifest records attempted 98, already_done 12: 12 rows survived from the aborted launches and the run resumed. Cost: about 20 minutes and a reorder of the cases for prompt-cache reuse (which then did not work: every case prefilled 9,145 tokens in 22.3 s).
Benchmarking on the live driver. ornith-9b-kv-f16 and the synthesis/contract/epistemics batch ran against the driver seat while Hermes used it, and the q4_0 arm ran on a separate port. The assistant gated the other tests on the driver being idle and the arms ran sequentially, but the f16 arm and the 16:31Z batch were not isolated from Hermes traffic. Median latency was 24.4 s vs 24.5 s between arms.
Different KV arms, different contexts. The f16 arm served 131,072 ctx; the q4_0 arm 49,152. Not an error (both exceed the 32k cases) but the arms do not share a server configuration beyond KV type.
Driver seat mislabelled on 2026-09-13 20:47Z. The alias ornith-1.5-9b served Qwen3.8-27B for the rest of that day. No run-log row falls in the window.
Stale or missing build stamps. Run-log llama_build for every row is 0.4.0-dev (build 10893, commit c32d1dabe), matching the manifests' /props (b10893-c32d1dabe); no mismatch. server_flags is null for all rows, so KV type is recoverable only from the commands above.
--jinja: on for every label. Reasoning-format flag absent, with thinking left at the template default for the 09-12 runs; the extraction profile handles it.

What happened to it

Downloaded 2026-09-12 ~09:21Z; benched the same day (09:3x-23:28Z); seated as Hermes driver 10:08Z; displaced 2026-09-13 ~20:12Z; swap-seat occupant 2026-09-28 onward.
On disk (ls -l models/ornith-1.5-9b, 2026-10-23): Ornith-1.5-9B-Q4_0.gguf 5,482,832,768 B and the mmproj 918,165,728 B. Not touched by any sweep: the 2026-10-11 safetensors sweep notes the "swap seat (ornith-9b) untouched" (a lab note), and RESTORE-deleted-20261011-tiersAB.sh has no entry for it.
Re-downloadable from bartowski/Ornith-1.5-9B-GGUF or ornith-ai/Ornith-1.5-9B-GGUF.