vibehacker
Discuss
Elena
10 hours ago

ollama swapped my 32b coder for a 3b mid-refactor

#local-llm#vibe-coding#rants

Left ollama running qwen2.5-coder:32b overnight on a Mac mini while Cursor agent chewed through a refactors queue.

Woke up to a PR that looked fine until review. Half the diffs were the kind of vague renames a tiny model loves. Checked ollama ps — somehow it was serving qwen2.5-coder:3b. Same name prefix. No crash. No alert.

I never pinned the tag in the Continue config. That's on me.

Do people hard-pin model digests for local agent runs, or is "whatever ollama has loaded" still normal?

5

5 comments

Join the discussion

Log in to comment.

Log in
  • Mira Chen

    Same class of bug bit me last week on a Figma-to-code pass.

    I asked for layout tokens. Got back button labels like "BtnPrimaryV2_temp". Opened the agent log — model line said qwen2.5-coder:1.5b. I had never typed that.

    Pinned digests in Continue after that. Prefix matching is a footgun when you also keep the small variants around "for speed".

    • Hazel Orbit

      the figma-to-code case is exactly why i stopped keeping small variants "for speed" on the same machine as the overnight agent.

      mobile side is worse — flutter cold start already eats 2s, then a 1.5b sneak-in makes the spinner look hung while it invents layout tokens. i pin digests in ollama Modelfile + a tiny preflight that fails the run if ollama ps != expected sha.

      prefix matching feels convenient until friday. then it costs you a weekend.

  • Priya Nair

    yeah i log ollama ps into the agent transcript now before every run. felt paranoid until a friday night "stabilize types" patch landed with the wrong model.

    hard-pinning the digest is the floor. also: don't keep :3b / :1.5b variants pulled if your config only says qwen2.5-coder — ollama will happily resolve to whatever fits RAM.

    do you pin in Continue, or in the Cursor model dropdown too?

    • Maple Drift

      Continue only, for me. Cursor's model dropdown still points at whatever cloud slot I last used — I don't trust it for local pins.

      What actually bit me: LiteLLM had qwen2.5-coder as an alias that resolved to whichever ollama tag was warm. Spreadsheet said 32b. Runtime said 3b. I now force model: ollama/qwen2.5-coder:32b@sha256:... in the router config and refuse short names in CI.

      Prefix matching is fine for humans. Terrible for overnight queues.

  • Spruce Glyph

    hard-pin the digest. period.

    i keep a one-liner in ~/.continue/config.json with the full sha, not the tag. also deleted every :1.5b / :3b pull so ollama can't "helpfully" resolve the short name when RAM dips overnight.

    if your agent can start a run without printing ollama show --modelfile into the log, you're gambling.

More like this

View all