expectedwrong hindsight

GPT Was Down So I Tried the Other Thing

Mistral dropped Codestral this morning and it's the first code model that made me forget OpenAI was having an outage.

2 min read 358 words #models #mistral #local-inference #ollama #coding
hindsight — nailed it

local models as cloud fallback became a standard pattern. 'ollama pull' as the response to 'GPT is down' is now normal. codestral proved the thesis and mistral kept shipping code-specific models. the laptop fallback is permanent infrastructure now.

GPT was down this morning — not degraded, not "some users may experience issues," just down — and rather than sit there refreshing the status page like a person whose plans have been cancelled by weather, I pulled Codestral.

ollama pull codestral. That's it. That's the install.

Mistral dropped it today. 22 billion parameters, 32k context window, trained specifically for code — not a general model that has been told to try harder on programming problems, but a model built from the floor up for the thing. It knows 80+ languages, which sounds like a marketing number until you actually try the obscure ones and they work.

The benchmark numbers are fine. They're always fine. What I care about is whether the model writes code that runs, and Codestral writes code that runs. It completes functions the way you meant them, not the way you literally described them. There's a difference, and most models land on the wrong side of it more than you'd expect.

The thing that gets me is the ollama path. You don't need an API key, you don't need to think about tokens, you don't need to care that Mistral's servers are up. It's on your machine. It runs while your internet is out. It ran this morning while OpenAI was busy having a moment.

The forced substitution is doing something real here. If GPT had been up I never would have pulled this. I would have filed it under "interesting, will evaluate later," which is the folder where models go to die. Instead I spent two hours with it on actual work and now it's in my rotation.

Mistral is a weird company — French, fast, genuinely technical, committed to the kind of open weights philosophy that the big labs keep saying they believe in right up until it becomes inconvenient. Codestral ships under a non-commercial license, which is not fully open, which is fine, which is also slightly annoying, which is where we live now.

But you can run it locally right now with one command. The model is excellent. OpenAI's infrastructure having a bad morning turned out to be useful information.