DGX model catalog

Authoritative list of Ollama models pinned on the DGX Spark, with their manifest digests. Per RFC-089 decision #7 — embedding model addition per ADR-098 (supersedes RFC-089 §D4).

How to read this file

• Tag is what the pipeline asks Ollama for (vector_embedding_model, ollama_summary_model, etc.).
• Digest is Ollama's manifest SHA — pin this to detect upstream model drift.
• Pulled is the UTC date the digest was captured on the operator's DGX.

How to refresh a row

After ollama pull <tag> on DGX:

ssh <dgx-host> "ollama show <tag> --modelfile | grep -E '^FROM' "

Or read the manifest directly:

ssh <dgx-host> "cat ~/.ollama/models/manifests/registry.ollama.ai/library/<tag-without-version>/<version>"

Update the row + date below; commit.

Catalog

LLM (chat / completion)

Digest = the weights-layer SHA from each manifest (application/vnd.ollama.image.model). Refresh with sudo jq -r '.layers[] | select(.mediaType == "application/vnd.ollama.image.model") | .digest' /usr/share/ollama/.ollama/models/manifests/registry.ollama.ai/library/<tag>.

Tag	Purpose	Pulled
llama3.3:70b	General autoresearch / GI/KG	2026-06-06
qwen2.5:72b-instruct	Comparison runs	2026-06-06
gemma2:27b	Comparison runs	2026-06-06
gpt-oss:120b	Operator-pulled (not RFC-089)	2026-06-05
qwen3.5:35b	Prod LLM champion (kept by #924 — see eval report linked below)	2026-06-08
qwen3.5:27b	#928 championship candidate (highest RougeL; latency-disqualified for prod)	2026-06-08
qwen3.5:9b	Comparison cell (#912 bundled-JSON flakiness known)	2026-06-08
qwen3.6:latest	Champion contender via #924 — pending #932/#933 validation before swap	2026-06-08
qwen3-coder:30b	Operator personal track (excluded from autoresearch by design)	2026-06-08
gpt-oss:20b	#924 sweep — closest non-Qwen3 to baseline	2026-06-08
deepseek-r1:7b	#924 sweep — R1 distill (reasoning-tuned, not summary-shaped)	2026-06-08
deepseek-r1:14b	#924 sweep — best of R1 distill family, still well below baseline	2026-06-08
deepseek-r1:32b	#924 sweep — slower AND worse than 14b	2026-06-08
deepseek-r1:70b	#924 sweep — operationally disqualified (~8 min/ep; killed mid-rerun)	2026-06-08
gemma3:27b	v2.1 sweep — 0.207 RougeL (-21%); likely prompt-template issue	2026-06-09
phi4:14b	v2.1 sweep — 0.256 RougeL; best 14B-class but latency-disqualified	2026-06-09
hermes3:8b	v2.1 sweep — 0.218 RougeL; Nous fine-tune regression vs base llama3.1	2026-06-09
mistral-small:24b	v2.1 sweep — 0.257 RougeL; strong mid-tier but 29s latency, not prod-viable	2026-06-09

Full digests:

• llama3.3:70b → sha256:4824460d29f2058aaf6e1118a63a7a197a09bed509f0e7d4e2efb1ee273b447d
• qwen2.5:72b-instruct → sha256:6e7fdda508e91cb0f63de5c15ff79ac63a1584ccafd751c07ca12b7f442101b8
• gemma2:27b → sha256:d7e4b00a7d7a8d03d4eed9b0f3f61a427e9f0fc5dea6aeb414e41dee23dc8ecc
• gpt-oss:120b → sha256:6be6d66a3f546d8c19b130dc41dc24b2fc159f84ffbc76a0ee0676205083cf5a
• qwen3.5:35b → sha256:900dde62fb7ebe8a5a25e35d5b7633f403f226a310965fed51d50f5238ba145a
• qwen3.5:27b → sha256:d4b8b4f4c350f5d322dc8235175eeae02d32c6f3fd70bdb9ea481e3abb7d7fc4
• qwen3.5:9b → sha256:dec52a44569a2a25341c4e4d3fee25846eed4f6f0b936278e3a3c900bb99d37c
• qwen3.6:latest → sha256:f5ee307a2982106a6eb82b62b2c00b575c9072145a759ae4660378acda8dcf2d
• gpt-oss:20b → sha256:e7b273f9636059a689e3ddcab3716e4f65abe0143ac978e46673ad0e52d09efb
• deepseek-r1:7b → sha256:96c415656d377afbff962f6cdb2394ab092ccbcbaab4b82525bc4ca800fe8a49
• deepseek-r1:14b → sha256:6e9f90f02bb3b39b59e81916e8cfce9deb45aeaeb9a54a5be4414486b907dc1e
• deepseek-r1:32b → sha256:6150cb382311b69f09cc0f9a1b69fc029cbd742b66bb8ec531aa5ecf5c613e93
• deepseek-r1:70b → sha256:4cd576d9aa16961244012223abf01445567b061f1814b57dfef699e4cf8df339
• gemma3:27b → sha256:a418f5838eaf (full digest captured 2026-06-09)
• phi4:14b → sha256:ac896e5b8b34 (full digest captured 2026-06-09)
• hermes3:8b → sha256:4f6b83f30b62 (full digest captured 2026-06-09)
• mistral-small:24b → captured on next dgx-deploy refresh

RFC-089 originally listed gemma2:27b-instruct and qwen2.5:72b-instruct. The -instruct suffix on gemma2 is not a real Ollama tag — gemma2:27b IS the instruct-tuned default; on qwen2.5 the suffix is correct. Catalog reflects what was actually pulled.

The autoresearch matrix (qwen3.x family, gpt-oss:20b, deepseek-r1 distills, qwen3-coder) was pulled across 2026-06-08 to support #924's smoke v2 refresh sweep. The v2.1 sweep added gemma3:27b, phi4:14b, hermes3:8b, and mistral-small:24b on 2026-06-09; results in the same eval report (Addendum section). None produced a new champion contender; qwen3.5:35b stays prod and qwen3.6:latest remains the only validated challenger pending #932/#933.

Embeddings — not used on DGX by default (ADR-098)

DGX does not serve embeddings in shipped profiles. The pipeline uses sentence-transformers/all-MiniLM-L6-v2 in-process on the host. The A/B in data/eval/embedding_provider_comparison/transcript-chunked/ showed MiniLM beats nomic under production-realistic chunking on this corpus.

Tag	Purpose	Dim	Context	Digest (weights layer)	Pulled
nomic-embed-text	Optional — for A/B eval reruns	768	8192	sha256:970aa74c0a90 (274 MB)	2026-06-06

Full weights digest: sha256:970aa74c0a90ef7482477cf803618e776e173c007bf957f635f1015bfcfef0e6 Captured from laptop Ollama 0.19.0; verify after ollama pull on DGX with cat ~/.ollama/models/manifests/registry.ollama.ai/library/nomic-embed-text/latest and check the layer of type application/vnd.ollama.image.model.

Pull only if you want to re-run the embedding-provider A/B against a newer model release. Not required for the pipeline.

Speech (Whisper)

Deferred to #814 — Whisper Large v3 is not in the Ollama library; the mechanism choice (Ollama vision/audio extension vs separate FastAPI shim vs alternative) is the prod-Whisper-via-DGX phase's decision. Pinning waits until that decision lands.

Drift policy

Any new pull or version bump on a pinned row → update this file in the same PR that bumps the profile or code referencing the tag. CI doesn't enforce this yet; reviewer responsibility.