Use Voxtral-Mini-3B-2507-bf16 across all tiers and set appropriate max_tokens for full-file transcription