Devstral-Small-2-24B lands Opus reasoning tune

This is less a SOTA chase than a behavior-packaging exercise: the value is in getting more deliberate coding out of a 24B model people can actually run locally.

• –The ~2.3k-sample dataset is small, so expect a planning/style lift rather than a dramatic jump in raw capability.
• –The VLM-to-text-only extraction is the real feat; it makes a multimodal base trainable on a single RTX 3090.
• –GGUFs plus a LoRA adapter are the right packaging for this audience: easy local testing, easy merges, easy Q4 vs Q5 comparisons.
• –The epoch-2 checkpoint choice, 2,048-token cap, and no-benchmark caveat suggest the model is tuned for practical generalization, not leaderboard theater.
• –Apache 2.0 licensing and open weights make it a useful sandbox for local coding agents and reasoning distills.