Devstral Small 2 tops local codebench

Interesting signal, but still a home-built benchmark, so the score matters more as directional evidence than as a universal ranking. The bigger takeaway is that Mistral may have finally shipped a local coding model that feels competitive for agentic code work instead of merely “good for local.”

• –Mistral’s own docs position Devstral Small 2 as an open-source model for tool use, multi-file edits, and codebase exploration, with a 256k context window and local deployment support.
• –The official launch claims 68.0% on SWE-bench Verified, so the Redditor’s stronger private benchmark result is plausible, but likely sensitive to task mix, prompts, and quantization.
• –The Q8 Unsloth GGUF is probably doing real work here; local coding performance often swings on quant quality, tool-calling behavior, and template correctness as much as raw parameter count.
• –Slow TPS is the expected tradeoff for a 24B local model, but if wall time stays acceptable and the fixes are high quality, it becomes a serious offline/private coding option.
• –Scaffold-style benchmarks can reward the exact behaviors they measure and miss regression risk, code style, and maintainability, so production testing still matters.