v0.3.0 — Devstral Small 2 Support
What's New
Devstral Small 2 24B (Mistral)
First Mistral family model added to the lineup. A dense 24B coding agent model purpose-built for agentic tool calling.
- 100% first-attempt tool call accuracy in testing — zero retries, zero validation failures
- Maintained accuracy across 40+ message conversations with 10+ chained tool calls
- 1-7s per tool call on A30 GPU (15GB)
- Apache 2.0 license, 68% SWE-Bench Verified
- Requires
num_ctx = 65536to prevent task wandering on open-ended searches
Other Changes
- Version bump to 0.3.0
- Profile tests for Devstral (pattern matching, settings, escalation)
- Updated deployment docs, hardware tables, and README
Full Changelog: v0.2.0...v0.3.0