feat(mistral-ai): update model YAMLs [bot]

[harshiv-26]

Auto-generated by poc-agent for provider mistral-ai.

---

Note

Low Risk
Low risk config-only changes, but they can affect model selection/capability gating (e.g., json_output, modality support, token limits) for Mistral models if downstream consumers rely on these fields.

Overview
Refreshes Mistral AI model definitions to align with current capabilities and lifecycle status by adding status (active/preview) across many YAMLs.

Updates capability metadata: adds json_output to several chat models, adds/adjusts modality support (e.g., embeddings now declare output: [embedding], mistral-ocr-latest adds text input, voxtral-mini-2507 adds image input, and some models drop previously listed image/audio inputs).

Adjusts limits/params where needed, including embedding output_vector_size, mistral-medium-latest max output/token caps and audio output, and introduces a default reasoning_effort param for magistral-medium-2509.

^{Written by Cursor Bugbot for commit e169097. This will update automatically on new commits. Configure here.}

[cursor]

Cursor Bugbot has reviewed your changes and found 3 potential issues.

^{Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

[cursor]

Devstral 2 incorrectly declares image input support

High Severity

image was incorrectly added to the input modalities for Devstral 2. According to Mistral's own documentation, Devstral 2 is explicitly text-only: "Text-only input — image or audio workflows require a separate model in the pipeline." This affects both devstral-2512.yaml and devstral-latest.yaml, and will cause consumers to incorrectly believe these models accept image inputs.

Additional Locations (1)

• providers/mistral-ai/devstral-latest.yaml#L14-L15

 

[cursor]

Voxtral Mini incorrectly declares image input support

High Severity

image was incorrectly added to the input modalities for Voxtral Mini 2507. Voxtral Mini is an audio-focused model for transcription and voice processing, and Mistral's documentation does not list vision/image as a supported capability. Notably, the corresponding voxtral-mini-latest.yaml correctly removes image in this same PR, making these changes contradictory.

 

[cursor]

Mistral Large incorrectly given thinking/reasoning capability

High Severity

thinking: true is incorrectly added to mistral-large-latest. Every other model with thinking: true is either a Magistral model (purpose-built for reasoning) or a Mistral Small 3.2+/4 model (with a reasoning_effort param and reasoning docs source). mistral-large-latest has none of these: no reasoning_effort param, no reasoning source URL, and its pinned counterpart mistral-large-2512 does not have thinking: true despite sharing the same docs URL. Web search confirms Mistral Large is separate from reasoning-focused model lines.