Thanks for the detailed report, including the screenshot. I reproduced this and there is an easy workaround while I fix it properly.
The cause: LFM2 is not a reasoning model, and the AI Assistant asks Ollama for thinking mode by default. Ollama rejects that request outright with a "model does not support thinking" error, but Studio currently swallows the real error and shows the generic "provider stream failed before completion" message you saw. The validation in Settings passes because it only asks Ollama for its model list, which never requests thinking.
To fix it right now: open Settings, go to the AI Assistant page, and untick "Enable thinking/reasoning". Chat should work immediately with LFM2. Alternatively, switching to a model that supports thinking (qwen3, deepseek-r1, gpt-oss and similar) works with thinking left on.
I have logged the bug here: https://feedback.objo.dev/bug/1826. The fix will stop Studio requesting thinking for models that don't support it and will surface the real provider error instead of the generic one.
