Only three AI models have been tested - help wanted trying the others #38

Clips Studio will drive any model Ollama serves, and the Models page lists several. Only three have actually been run against real streams:

• gemma:7b
• gemma3:4b
• gemma3:12b

Everything else — Gemma 4 (e2b, e4b, 12b), Qwen3, Mistral Nemo, Phi-4, Llama 3.1, Qwen2.5 — is listed because it is a sensible size for the hardware and free to use commercially, not because clip quality has been measured with it.

They should all work; the app talks to them identically through Ollama. What nobody knows is whether they pick better or worse moments, which is the thing that actually matters.

How to help

Run one on a video you know well and say what happened. Useful to include:

• which model and size
• what kind of content (just chatting, podcast, IRL, vlog)
• whether the clips it chose were better, worse or indistinguishable from Gemma 3
• anything it got obviously wrong — bad start or end points, missed the funny bit, picked dead air

No formal benchmark needed. Several people saying "qwen3:8b picks better podcast moments" is worth more than one synthetic score.

Particularly wanted:

• Gemma 4 e2b and e4b — these are recommended for low-power machines, so they would be many people's first experience of the app.
• Qwen3 for non-English content — it is listed as the translation recommendation on reputation rather than measurement.

Switching model is one line: pull it in Ollama, then pick it on the Models page.