Ollama Sync

Compare the :cloud models listed on ollama.com against the models installed locally on this PC, generate a run script for the web-only models, and test which installed models actually work — with colour-coded ✓/✗/⚠ results in the console and a timestamped log of every run.

Files

File	Purpose
ollama_sync.py	Main tool: manage Ollama, scrape web :cloud models, diff vs local, write log + run script.
test_models.py	Test every installed model with a hi prompt and classify it (functional / needs subscription / retired / not working / unavailable). Colored ✓/✗/⚠ output on the console.
run_sync.bat	Launcher for ollama_sync.py — runs it, lists the log files, and keeps the window open with a key-press.
test_models.bat	Launcher for test_models.py — runs it, lists the log files, and keeps the window open with a key-press.
run_cloud_models.bat	Generated — one ollama run <model> (or ollama pull with --pull-only) per web-only model.
outputs/web_cloud_models.txt	Generated — deduplicated sorted list of web :cloud models.
outputs/local_models.txt	Generated — sorted list of local models from ollama list.
logs/ollama_sync_<timestamp>.log	Generated — every sync run writes a log here.
logs/model_test_<timestamp>.log	Generated — every model test run writes a log here.

How the sync works

1. Manage Ollama — checks tasklist for ollama.exe / ollama app.exe. If not running, starts ollama serve and waits until http://localhost:11434/api/tags responds (up to ~30s). If this script started Ollama, it closes only that instance (by PID) again when done — other Ollama processes are left alone. With --no-close, even the instance this script started is left running. If Ollama was already running, it is left untouched. (Mirrors core_value_providers.manage_ollama.)
2. Read local models — queries the Ollama REST API (/api/tags) first (most reliable, real size/modified values), falling back to ollama list --format json and then plain ollama list text parsing.
3. Scrape web cloud models — fetches https://ollama.com/search?c=cloud&o=newest and walks the catalog page-by-page (p=1, 2, …) until two consecutive pages add no new models, collects the /library/<model> base models, then visits each model's /library/<model>/tags page and extracts every tag ending in cloud (e.g. glm-5.2:cloud, deepseek-v4-flash:0731-cloud, gemma4:31b-cloud). If any search or tags page fails, the whole web inventory is reported as FAILED (a partial crawl is not a complete inventory). (Based on unified_model_loading.py; uses the hidden <input class="command"> tag rows with a text-regex fallback, plus retries with exponential backoff and HTTP 429 handling.) No duplicates: tags are collected into a set, cross-model duplicates are reported and kept once, and the final list is case-insensitively deduplicated.
4. Show the difference — prints:
   • Extra on WEB — cloud models not installed locally.
   • Extra on LOCAL — local models not on the web cloud list.
   • Common — installed locally and on the web.
5. Write run_cloud_models.bat — one ollama run <model> line per extra-on-web model (ollama pull <model> with --pull-only). By default it first prints a PLANNED update block listing exactly which models will be added, then asks for confirmation ([y/N], Enter = No) before overwriting the file. Only a y/yes rewrites it; anything else (or non-interactive stdin) leaves the existing script untouched. --pull-only skips the prompt and regenerates directly, since that is an explicit request.
6. Write a timestamped run log to logs\ containing:
   • ALL IN WEB (deduplicated)
   • ALL IN LOCAL
   • NEW IN WEB (not installed locally)
   • NEW IN LOCAL (not on the web cloud list)
   • plus COMMON, an integrity summary (pages scanned, base models found, duplicates removed, failed pages, inventory status), summary counts, and a line recording whether run_cloud_models.bat was updated (yes/no).

If either inventory fails (network error, HTTP failure, a partial page crawl, or ollama list errors), that failure is reported in the console and log as FAILED — never silently treated as an empty list. --web-only records the local side as skipped (deliberately not checked), which is also distinct from empty.

Usage

run_sync.bat               REM run the sync, then shows the log file(s)
run_sync.bat --web-only    REM skip local Ollama management/list
run_sync.bat --pull-only   REM generate run_cloud_models.bat with `ollama pull`

Or directly: python ollama_sync.py [--web-only] [--no-close] [--pull-only]

• --web-only — skip the local Ollama open/list/close (local side is recorded as skipped — deliberately not checked — rather than empty).
• --no-close — keep Ollama running even if this script started it.
• --pull-only — generate run_cloud_models.bat using ollama pull (download only) instead of ollama run (interactive chat). Skips the confirmation prompt.

Without --pull-only, the script lists the web-only models it would add and asks Regenerate run_cloud_models.bat with these N web-only model(s)? [y/N] (Enter = No) — press y to overwrite, or just press Enter / anything else to leave the existing script as-is.

Testing models

test_models.py sends a hi prompt to every installed model via the Ollama API and classifies the result. Each model gets a generous first attempt (300 s) so a slow cold-start isn't misclassified as dead; if that times out, it retries once with a shorter 120 s timeout. The reported elapsed time covers both attempts.

Status	Symbol / colour	Meaning
functional	✓ green	Responded normally to hi. Ready to use.
functional (slow)	✓ cyan	Responded, but took > 60s.
needs subscription	⚠ yellow	Model exists but requires an Ollama subscription/upgrade (HTTP 401/403). Not usable without upgrading.
retired (410 gone)	✗ gray	Model is no longer served by Ollama's cloud (HTTP 410). Not usable.
not working (no response)	✗ red	Timed out or returned an empty reply.
not working (error)	✗ red	Request failed for another reason.
unavailable (not installed)	– magenta	Requested via --only but not present locally.

The console output is colour-coded (Windows 10+ cmd): the per-model progress line shows the symbol + status coloured by result, and the summary counts and grouped lists use the same colours. Log files are written as plain text (no colour codes) but mirror the full report: per-model results (status, elapsed, reply snippet, error), the grouped model lists, and the summary counts.

test_models.bat                 REM test every installed model
test_models.bat --limit 5       REM only the first 5 models
test_models.bat --only glm-5.2:cloud   REM test a single model
test_models.bat --parallel 3    REM test 3 models concurrently (watch VRAM)

--parallel above 4 prints a VRAM warning — concurrent large models can exhaust GPU memory. If the local model inventory cannot be read (ollama list fails), the test script aborts with a clear error and a non-zero exit code instead of pretending there are zero models.

Each run writes a per-model log under logs\. When launched via the .bat, the window stays open until a key is pressed so the report can be read on screen.

Requirements

• Windows 10/11 (the lifecycle helpers use tasklist / taskkill / CREATE_NO_WINDOW; Linux/macOS would need process-control changes)
• Python 3.x
• requests, beautifulsoup4 (pip install requests beautifulsoup4)
• Ollama CLI on PATH

---

C:\Users\ADMIN>ollama list NAME ID SIZE MODIFIED kimi-k3:cloud 630e737485bd - 16 hours ago gemma4:cloud ef09f235533c - 16 hours ago deepseek-v4-flash:preview-cloud 5166728b9358 - 16 hours ago deepseek-v4-flash:0731-cloud d3f1c8744721 - 16 hours ago glm-5.2:cloud ce8fd6f94793 - 6 weeks ago kimi-k2.7-code:cloud eda07a659237 - 8 weeks ago nemotron-3-ultra:cloud 6d55374b63bb - 2 months ago minimax-m3:cloud d03a959f45c0 - 2 months ago deepseek-v4-pro:cloud 22bfd5026abd - 3 months ago deepseek-v4-flash:cloud ea027821675c - 3 months ago kimi-k2.6:cloud a90cd0d1590c - 3 months ago glm-5.1:cloud 59472abf9d0a - 3 months ago gemma4:31b-cloud c382fbfbc73b - 4 months ago minimax-m2.7:cloud 06daa293c105 - 4 months ago nemotron-3-super:cloud be3943c5a818 - 4 months ago qwen3.5:397b-cloud a7bf6f7891c3 - 5 months ago qwen3.5:cloud a7bf6f7891c3 - 5 months ago minimax-m2.5:cloud c0d5751c800f - 5 months ago glm-5:cloud c313cd065935 - 5 months ago qwen3-coder-next:cloud aa626c11ae8d - 6 months ago kimi-k2.5:cloud 6d1c3246c608 - 6 months ago glm-4.7:cloud 023608864819 - 7 months ago minimax-m2.1:cloud 4ada3a038304 - 7 months ago gemini-3-flash-preview:latest ebade0d31690 - 7 months ago gemini-3-flash-preview:cloud 436200142af2 - 7 months ago nemotron-3-nano:30b-cloud 01d0d069a149 - 7 months ago gemini-3-pro-preview:latest 91a1db042ba1 - 7 months ago rnj-1:8b-cloud d8200a2fbf21 - 7 months ago devstral-small-2:24b-cloud ec4a591da58a - 7 months ago devstral-2:123b-cloud d37aca5b6a27 - 7 months ago qwen3-next:80b-cloud f5ccd68d2872 - 7 months ago gemma3:27b-cloud 9e1580299085 - 7 months ago gemma3:12b-cloud 485e7119a53a - 7 months ago gemma3:4b-cloud 89c58fea5420 - 7 months ago deepseek-v3.2:cloud 55f7c48fb187 - 7 months ago mistral-large-3:675b-cloud 3130fd5a5a1e - 7 months ago ministral-3:14b-cloud 615c59440878 - 7 months ago ministral-3:8b-cloud a56a5396dfb9 - 7 months ago ministral-3:3b-cloud 6938c17dead4 - 7 months ago cogito-2.1:671b-cloud 36c90b0682ed - 7 months ago qwen3-vl:235b-instruct-cloud 2bf9522f6961 - 7 months ago deepseek-v3.1:671b-cloud d3749919e45f - 7 months ago gpt-oss:20b-cloud 875e8e3a629a - 7 months ago glm-4.6:cloud 05277b76269f - 7 months ago kimi-k2-thinking:cloud 9752ffb77f53 - 7 months ago minimax-m2:cloud 698ab6d56142 - 7 months ago qwen3-coder:480b-cloud e30e45586389 - 7 months ago kimi-k2:1t-cloud 20dc43ca06d7 - 7 months ago qwen3-vl:235b-cloud 86b3322ec200 - 7 months ago gpt-oss:120b-cloud 569662207105 - 7 months ago

C:\Users\ADMIN>

---