FreshContextA context lab · Fife
Drawing
02 · Benchmarks
Sheet
2 of 3
Rev
A · 2026-10-01

Ornith 1.5 35B-A3B · file← Ornith 1.5 35B-A3B

Q6_K

Our everyday model in late August, and the first one scored. Deleted to free disk.

Size
29.2 GB
Status
Deleted 8 Sep 2026
Scored runs
3
Smallest Mac
48 GB estimate, 8K context
Results3 scored · 95% intervals
SuiteScore95% interval PassedRuntime · profileRecorded
groundingornith-1.5-35b-q6-b1061872.8%64.4–79.8Passed 91/125llama.cppextraction2026-08-25
nullschemaornith-1.5-35b-q6-b1061834.7%24.8–46.2Passed 25/72llama.cppextraction2026-08-25
screeningornith-1.5-35b-q6-b1061895.0%86.3–98.3Passed 57/60llama.cppextraction2026-08-25

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.

FileOrnith-1.5-35B-Q6_K.gguf (+ mmproj-Ornith-1.5-35B-BF16.gguf, 0.90 GB)
FormatQ6_K, 29.21 GB
Where fromdownloaded 2026-08-21 (ornith-q6-dl.log); repo not in the log
Statusdeleted 2026-09-08 16:16Z
Hugging Faceas above (filename + size match)

Where it came from

  • Repo: inferred, not logged: ornith-ai/Ornith-1.5-35B-A3B-GGUF by exact filename and size (listing).
  • Revision, sha256: not on record. The harness manifest's partial hash is in benchmarks/results/*__ornith-1.5-35b-q6-b10618__*.manifest.json only if weights was recorded; the 08-25 manifests carry no server block.
  • Date: 2026-08-21 (UTC 2026-08-21 00:11 by the log mtime). Who asked: not on record.

How it was served

labeldates (UTC)binary and versionctx / -np / KV / -fa / --jinja / --reasoning-formatharness profilewhy this label exists
ornith-1.5-35b-q6-b106182026-08-25 ~08:59–09:03Z (run-log ts 10:00:41–10:03:30 BST)llama.cpp/build/bin/llama-server, b10618 (eb25b7263)-c 32768 -np 4 -cb (8,192 per slot) / f16 KV / -fa not set (default) / --jinja yes / reasoning-format default; no mmprojextraction (thinking off, max_tokens 2048, temperature 0)the locked v2 suites' first run of this model, after the b10618 rebuild fixed tool-call JSON; -b10618 suffix keeps it separate from build-10369 rows

Command: a one-off script, harness.runner --suite {grounding,screening,nullschema} --model-id ${n}-b10618 --profile extraction --no-resume, after a tool-call gate and a Toolery run on the same server.

Seats held (off the run log): router preset [ornith-1.5-35b] with c = 262144, q8_0 KV, reasoning-budget = 8000, mmproj, load-on-startup (fleet-models.ini); resident default and production enrichment from 2026-08-21 (a lab note). Load-on-startup stood down 2026-08-26; preset pruned 2026-09-08.

Decisions

  • Downloaded and made default 2026-08-21 over Q5_K_M (fleet-models.ini).
  • Measured 2026-08-25 on Alan's "we never put ornith through the gauntlet".
  • Kept 2026-08-26 "for coding".
  • Deleted 2026-09-08 16:16Z. Alan: "nuke ornith for now, we have more interesting models on the go"; the assistant had listed it as a disk candidate: "coder scored only on extraction, beaten by devstral". Twelve result files kept.
  • Not published to HF (downloaded file).

What went wrong

  • First attempt void (2026-08-24): two model_run.sh jobs loaded a 27 GB and an 18 GB model at once; the readiness loop had no failure branch and the gate reported "template broken". Then b10369 emitted malformed tool-call arguments for this model; fixed by the b10618 rebuild. A zsh print mangling \" made the fixed build look broken for half an hour.
  • Duplicate presets crash (2026-08-21; see the Q5_K_M doc).
  • The "15/15" record was never revised after the locked suites scored this file at nullschema 34.7.

What happened to it

Downloaded 2026-08-21 → resident default → locked suites 2026-08-25 → preload stood down 2026-08-26 → deleted 2026-09-08 16:16Z with its mmproj; preset commented "pruned 2026-09-08: weights deleted" (fleet-models.ini). No RESTORE entry. Re-downloadable from https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B-GGUF (Ornith-1.5-35B-Q6_K.gguf).