rdmsm4x-changelog-20260922-1850-ollama-model-refresh

Loaded three exact local open-weight Ollama model tags additively on the canonical Studio, benchmarked synthetic receipts, and preserved all existing routes and models.

Scope and reason

• Host changed: rdmsm4x (Production Ollama model store and documentation). Read-only checks on five peers; separate host-local changelogs cover the two Intel canaries.
• Richh requested refreshed trusted local models for OCR/vision, fast work, and coding. No app, service, client, credential, or network configuration was changed.

Files and state changed

• Ollama model store: /Users/richh/.ollama received only official-registry glm-ocr:q8_0, qwen3.5:4b-mlx, and qwen3-coder-next:q4_K_M manifests/blobs. All ten pre-existing Studio tags were retained. Store size after: 367G by du -sh.
• Project: /Users/richh/dev/localAI/ollama/AGENTS.md, README.md, SESSION-STATE.md, ISSUES.md, docs/model-selection-and-fleet-rollout-2026-09-22.md, and scripts/benchmark_ollama_receipts_v1.0.py.
• Coordination: /Users/richh/.agent-coordination/checkins/codex-ollama-fleet-setup-20260826.json was refreshed and synced to all reachable hosts. Canonical iCloud PROJECTS.md received one Ollama entry. A bounded research worker wrote a non-secret source report at /Users/richh/dev/_handoff/codex-out/ollama-model-research-20260922.md and its own completed check-in.
• No Git repository was initialized in the Ollama project root. No existing model was removed, retagged, overwritten, or routed to a client.

Commands and verification

• ollama pull glm-ocr:q8_0, ollama pull qwen3.5:4b-mlx, ollama pull qwen3-coder-next:q4_K_M all exited successfully with digest verification and manifest write.
• curl http://127.0.0.1:11434/api/tags showed 13 tags and full new digests: 2a5a0f1a93017fc9db321ec196efb4b9bbba97c4d890df8e39429ed771f2ed25, 61aa3858e9d3022e8fca725550089addd9289c4446f33bd09dfd12f95f2a6792, ca06e9e4087c714d44355bf954099187890e63084b4a632b8e9956c4b9492074, respectively.
• uv run --script scripts/benchmark_ollama_receipts_v1.0.py scored each of GLM-OCR, Qwen 3.5 4B, Qwen 3.8 27B, and Gemma 4 31B at 50/54 synthetic receipt fields/items. Median warm times were 1.93, 2.19, 5.19, and 4.95 seconds. GLM-OCR hit the 512-token cap on all five, so it was not promoted as an unattended default.
• The same script with --suite photos scored two simple synthetic drawings on scene type, mug and lamp colors, and sign text: 4B and Qwen 3.8 each 8/8, Gemma 4 7/8. An ambiguous ground-truth laptop label was excluded; no real-photo accuracy claim is made.
• The script's --suite documents rendered page 1 of a synthetic two-page PDF into automatically cleaned temporary scratch. GLM-OCR, fast 4B, and Qwen 3.8 each found 10/10 title/heading/table markers; 4B completed in 4.22 seconds warm, while GLM-OCR again hit the output cap. The ad-hoc /tmp/ollama-doc-canary.2ZgAAc/page1.png rendering and its empty directory were removed after verification; the older unrelated temporary rendering was preserved. No original fixture or private document changed.
• A bounded Qwen3-Coder-Next code smoke completed in 16.08 seconds including 12.52 seconds load. ollama ps showed 51 GB, 100% GPU, 4K context; sysctl vm.swapusage showed 0 MB used and memory_pressure -Q reported 40% free. This proves compatibility, not coding-quality leadership.
• lsof and loopback API checks showed Ollama 0.34.2 still listening only on 127.0.0.1:11434. No service restart or route change occurred. Detailed reproducible evidence is in the project rollout document.

Backup, rollback, and outstanding gates

• Time Machine destination: smb://rich@unaspro818a/timemachine; latest reported Studio backup was 2026-09-22 14:22:59; tmutil isexcluded /Users/richh/.ollama reported Included. The backup's model-store contents and an isolated restore were not validated. No new backup or restore was claimed.
• Non-destructive rollback: leave the new tags unused, preserve their blobs/manifests for review, and retain every existing client route. Do not run ollama rm, manually delete blobs, or replace the Ollama app as a rollback shortcut. Any future deletion or route change needs a separate target-specific acceptance and recovery gate.
• Open: representative clean document/photo and coding benchmarks; backup/restore proof; M5 Air backup/storage gates; Intel account-lane topology anomalies. No owner physical or credential action was performed.