Devstral: the open coding agent model that earns its keep

Most open-weights models advertised for coding are chat models wearing a hoodie. They complete functions beautifully, then fall apart the moment an agent harness asks them to pick a tool, read the result, and decide what to do next. The loop is the hard part, not the syntax. Devstral is the exception I keep coming back to, because it was built the other way around: trained for the loop first, with an actual agent harness in mind from day one.

What the 25.12 revision changes

The original Devstral was a proof of concept with a 46 on SWE-bench Verified. Modest, but genuinely open and effectively free. The 25.12 revision is the grown-up version: 262k context, noticeably steadier tool-calling, and API pricing at $0.40 in, $2.00 out if you don't want to host it yourself. No published SWE-bench run for it yet, which annoys me and I've said so; my anecdotal runs put it well above its ancestor.

The context bump matters more than it sounds. Agentic coding eats context (file reads, tool results, failed attempts), and 128k was exactly the size where my sessions died mid-refactor. At 262k, the context budget stops being the first thing that kills a run.

The envelope

Here's the honest shape of what it can do, from a few weeks of running it behind Aider and OpenCode:

• Comfortably inside: bounded multi-file edits, test-driven fix loops, mechanical refactors, boilerplate with project conventions, review-and-patch cycles on diffs under a few hundred lines.
• The edge: long autonomous chains. Give it a vague ticket and twenty steps of freedom and it wanders, not catastrophically, just inefficiently, re-reading files and second-guessing working code.
• Outside: architectural judgement calls. It will confidently restructure something that didn't need restructuring. Keep a human, or a bigger model, on plan approval.

Devstral's job is to make 70% of your agent traffic stop needing the frontier model.

That framing is the whole economics. Routed sensibly (Devstral for the bounded work, a frontier model for planning and the gnarly 30%), my per-task costs dropped hard, same pattern as the local-first cascade. And because the weights are open, the floor price is your electricity bill: the 24B-class original ran on a single high-memory consumer GPU, and the new one is still in single-node territory, quantized.

Why open weights matter here specifically

Coding agents see your entire repo, your secrets hygiene, your internal APIs. Of all the places to want weights you control, this is the top of the list. It's the same argument I made in the sovereign stack piece, sharpened by the fact that Devstral is the first open model where the agentic behavior, not just the code knowledge, feels deliberate.

I wouldn't bet a product on it alone. The ceiling is real. But as the workhorse tier of a routed setup, it's the strongest open answer available right now, and it's not close. Start with your most mechanical ticket category, wire it into the harness you already use, and measure before you believe anyone, including me.