Mistral Vibe: Reasoning Effort Configuration and Per-Model Context Management
Mistral VibeMistral Vibe v2.4.0 introduced reasoning effort configuration alongside a thinking blocks adapter, giving developers direct control over how much compute reasoning models spend before responding in an agentic session. The release also made the auto-compact threshold per-model, allowing context compression behavior to be tuned independently for each model in a multi-model setup. Additional performance improvements — including cached git operations and elimination of expensive file scans from the system prompt — make agent switching noticeably faster, particularly in large repositories.
Reasoning Effort Configuration and Thinking Blocks Adapter
Mistral Vibe v2.4.0 added first-class support for reasoning effort configuration, giving developers direct control over how much compute a reasoning model expends before generating a final response. This is surfaced through a thinking blocks adapter, which translates model-level reasoning controls into the CLI's agent configuration layer, making the feature accessible without requiring manual API parameter tuning.
For developers using Mistral's reasoning-capable models, this means finer-grained control over the quality-versus-speed trade-off in agentic sessions. A higher reasoning effort setting encourages the model to reason more carefully before acting — particularly valuable for complex multi-file edits or architectural decisions — while a lower setting prioritizes responsiveness during exploratory or iterative sessions.
Per-Model Auto-Compact Threshold
Previously, Mistral Vibe applied a single auto-compact threshold across all models in a session. With v2.4.0, the auto-compact threshold is now per-model, meaning context compression behavior can be configured independently for each model in use. This is especially useful in mixed-model setups where a lightweight model handles fast tasks while a larger reasoning model handles deeper analysis — each can now manage its own context window compression independently, rather than sharing a single global limit that may be too aggressive for one model or too permissive for another.
Performance Improvements to Agent Switching
Two targeted optimizations reduce latency when switching between agents. Mistral Vibe now caches git operations that previously ran on every agent switch, and has removed an expensive file scan that was previously executed during system prompt construction. Combined, these changes make agent mode transitions faster, particularly in large repositories where filesystem traversal was the main bottleneck.
Plan Mode and CLI Banner
Mistral Vibe improved plan mode behavior in this release. The CLI banner now displays the user's current subscription plan, providing immediate visibility into account tier at startup without requiring a separate whoami command — a minor but practical quality-of-life addition for users on Le Chat plans.
Bug Fixes
- The space key stopped working in VSCode 1.110+ due to a terminal input handling change; this has been fixed.
- Arrow-key history navigation at wrapped-line boundaries in the chat input now behaves correctly.
- UTF-8 encoding is now enforced when reading metadata files, preventing corruption on systems with non-UTF-8 default locales.
- The update notifier no longer crashes when it encounters unexpected fields in the version check response.