Mistral Vibe: Thinking Levels Sent As Picked, Including Max on Mistral Large 4
Mistral VibeMistral Vibe 3.0.3 made thinking levels behave predictably. Levels are now sent to the model exactly as picked, max reasoning works on Mistral models that support it such as Mistral Large 4, and the xhigh level appears for models that list it. An unknown level no longer blocks startup, and picker options follow each model's configured thinking_levels.
Key Takeaways
- Thinking levels reach the model unchanged, so
maxis sent asmax. - Max reasoning effort now works on Mistral Large 4 and other supporting Mistral models.
- The
xhighlevel shows up for models that list it, and the picker follows each model'sthinking_levels. - An unset level uses the model's lowest, and
offis refused on always-reasoning models such as gpt-6-astra and GLM 5.3. - An invalid level no longer blocks startup; it is ignored with a warning.
- The
compaction_thinkingsetting was removed, a breaking config change for anyone who set it.
Reasoning effort without surprises
Mistral Vibe 3.0.3 cleans up how the thinking (reasoning effort) setting reaches the model. Previously, a chosen level could be silently changed on its way out. Now, levels are sent as picked: max goes as max and high as high.
What is new
- Reasoning effort
maxnow works on Mistral models that support it, such as Mistral Large 4. - The
xhighlevel is offered for models that list it. - When a model's config lists
thinking_levels, the picker offers those levels, plus the user's configured level if it falls outside them. - A level the model does not accept now fails with the provider's own error instead of being quietly adjusted.
Defaults and turning thinking off
An unset thinking value uses the model's lowest level. off is refused on models that always reason, such as gpt-6-astra and GLM 5.3. With third-party models that can turn thinking off, off does turn it off. Requests Vibe makes on its own, such as image descriptions and worktree names, follow the same rule: thinking is off when set to off, and models that cannot disable it use their lowest level.
Friendlier recovery
An unknown thinking or thinking_levels value no longer stops Vibe from starting. It is ignored with a warning, so a level can be chosen with /thinking. Switching from a reasoning model to a Mistral model without reasoning also no longer fails. Separately, the release removes the compaction_thinking setting, so compaction uses the compaction model's thinking level.