jdmbair13m5-changelog-20260922-1953-ollama-small-model-canary
jdmbair13m5-changelog-20260922-1953-ollama-small-model-canary
Two exact small Ollama models were added under a hard free-space floor for opt-in OCR/fast vision evaluation; both existing model tags and the loopback service were preserved.
- Scope:
jdmbair13m5only, operated from canonicalrdmsm4xover its pinned SSH identity. No service install, restart, client route, credential, account, or network change. - Files/state touched: only additive Ollama manifests/blobs under
/Users/richh/.ollamaforqwen3.5:4b-mlxandglm-ocr:q8_0; no project source file was written on this host. This sanitized changelog copy is archived at/Users/richh/dev/LLM/Claude/changelogs/jdmbair13m5-changelog-20260922-1953-ollama-small-model-canary.mdon the host and canonical Studio. - Commands: read-only
scutil,df, loopback/api/versionand/api/tags,ollama ps,lsof,tmutil, memory/swap checks; thenollama pull qwen3.5:4b-mlx, disk-floor recheck,ollama pull glm-ocr:q8_0; synthetic 4K receipt/photo evaluations through an SSH loopback forward;ollama stoponly those two canary-loaded tags; final inventory/listener/disk/swap checks. - Verification: Ollama 0.34.2 remained at
127.0.0.1:11434. New full digests were61aa3858e9d3022e8fca725550089addd9289c4446f33bd09dfd12f95f2a6792and2a5a0f1a93017fc9db321ec196efb4b9bbba97c4d890df8e39429ed771f2ed25; both old digests were unchanged. Qwen 4B matched 22/23 checked fields on two synthetic receipts and 8/8 simple-scene facts; GLM-OCR matched 22/23 but both outputs hit the 512-token cap. Both ran 100% GPU; swap stayed zero, and 33.1 GB (about 31.6 GiB) remained free, above the 25 GiB floor. No model remained resident after cleanup. - Backup:
.ollamawas marked included for Time Machine, but this host has no configured destination or isolated restore. No model-store backup or private key was copied. Existing models were not removed. - Undo/rollback: leave the new tags unused and retain their
manifests/blobs for review; existing routes remain unchanged. Do not run
ollama rmor manually delete blobs without a separately validated restore and exact deletion authorization. - Outstanding: low disk space, account-lane topology classification, real-photo/multi-page-document quality, OCR truncation handling, and an actual model-store restore remain open. The temporary synthetic-fixture SSH forward was stopped.