Mistral Devstral: Open-Source Coding Agents That Run on Your Laptop Just Changed the Game

Mistral Devstral: Open-Source Coding Agents That Run on Your Laptop Just Changed the Game

Here's something that didn't get enough attention when it happened, and it really should have. In December, French AI lab Mistral released Devstral 2, a pair of open-weight coding models that can do something none of the big players have pulled off: run a competent AI coding agent on consumer hardware, offline, for free.

The smaller version, Devstral Small 2, is 24 billion parameters and scores 68.0% on SWE-bench Verified. It fits on a single RTX 4090 or a MacBook with 32GB of RAM. The full 123B version scores 72.2%, which puts it within spitting distance of Claude Sonnet 4.5 on the same benchmark, and you can run that one on a workstation with four GPUs or a single H100. The code and weights are available under permissive licences (Apache 2.0 for Small 2, a modified MIT for the 123B model). No API key. No cloud dependency. No per-token billing.

I've been watching the local LLM scene for a while, and this is the first time I've thought: okay, this actually changes the calculus for real work. Not just for hobbyists. For professionals. And most interesting to me, for no-code and low-code builders who've been locked out of AI-assisted development because of data sensitivity rules, cost concerns, or just plain unreliable internet.

What Devstral actually is

Let me back up. Devstral isn't a general-purpose chatbot. It's a model trained specifically to work as a coding agent: it explores codebases, reads files, writes and edits code, runs commands, and iterates on solutions. Mistral built it in collaboration with All Hands AI, the team behind the OpenHands agent scaffold. The training data came from real GitHub issues, so the model learned to solve actual software engineering problems rather than just writing neat standalone functions.

The first Devstral dropped in May 2025 at 46.8% on SWE-bench Verified, which was already the best open-source score at the time. Devstral 2 in December pushed that to 72.2% for the full model and 68.0% for the small one. By May 2026 the models are stable, well-supported in Ollama (official library tags shipped), and integrated into all the major coding tools: Cline, Continue, Claude Code, OpenCode, and Mistral's own Vibe CLI.

What matters more than the benchmark number is what you can do with it. With Devstral Small 2 running locally via Ollama, you get sub-200ms first-token latency on a 4090. No network round trip. No queue. You can be on a plane, in a coffee shop with terrible WiFi, or in a building that blocks outgoing API calls to `api.anthropic.com`, and your coding agent keeps working.

Why no-code builders should care

Most people reading nocode.tech aren't spending their days writing Python in VS Code. You're building apps with Stacker, Bubble, Glide, or WeWeb. You're stitching together backends with Make.com or n8n. Your value is knowing what to build and how to connect things, not hand-optimising React components.

So why does a coding agent model matter to you?

Three reasons.

First, the privacy problem is real and it's blocking adoption. I've talked to no-code consultants who build internal tools for law firms, healthcare providers, and financial services companies. These clients have data that legally cannot leave their infrastructure. Every major vibe coding tool today (Cursor, Bolt, Lovable, Claude Code, GitHub Copilot) sends your code to a third-party API. That's a non-starter for regulated industries, full stop.

Devstral Small 2 running locally means you can vibe-code a client dashboard, an internal workflow tool, or a patient management interface without any of the underlying code, database schemas, or business logic touching a cloud endpoint. Your prompts, your chain of thought, your repository contents, all stay on the box. This is not a nice-to-have. For whole categories of software work, it's the difference between being allowed to use AI or not.

Second, the cost model flips. Devstral 2 on your own hardware has zero marginal cost after you've bought the GPU or Mac. Compare that to Claude Code, where a heavy agentic coding session can burn through $20-50 in API credits without blinking. If you're an indie dev or a small agency, that adds up fast. The API pricing for Devstral 2, when it exits preview, is expected at $0.40 per million input tokens and $2.00 per million output tokens. That's already cheaper than Claude, but local inference is free. There is no token counter ticking in the corner of your screen.

For no-code builders who use AI coding assistants to generate custom components, write API integrations, or debug JavaScript — not all day every day, but often enough — the difference between "I can use this whenever I want" and "I need to watch my monthly burn" changes how you work. You experiment more. You try things. You iterate faster.

Third, you don't need to be an engineer to benefit. The tooling around Devstral is maturing fast. Ollama's `ollama launch` command auto-wires local models into Claude Code, OpenCode, and Cline with a single terminal command. Mistral's own Vibe CLI is a terminal-native agent that works exactly like Claude Code except it's open source and speaks to your local Devstral instance. You type what you want, it reads your codebase and does it.

The bigger picture: open weights as business model pressure

Here's where I get opinionated. Devstral signals a future where the best coding agents aren't necessarily SaaS subscriptions. They're open-weight models you run yourself.

This is the same dynamic that's played out in other corners of software. Linux didn't kill Windows, but it created a permanent alternative that kept the proprietary OS vendors honest. Chromium didn't kill Safari, but it made sure no single company controlled how people browse the web.

I think Devstral does the same thing to the coding agent market. Cursor, Copilot, and Claude Code now have genuine open-weight competition that runs on consumer hardware. If Anthropic raises Claude's API prices too aggressively, there's a free alternative. If GitHub Copilot adds restrictive licensing terms, people have somewhere to go. Competition from below tends to compress margins, and compressed margins are good for developers.

The cloud-dependent tools aren't going anywhere. Claude Code with Sonnet 4.5 still scores higher on SWE-bench than Devstral 2 (roughly 39.8% vs 37.6% on one community comparison using the verified-mini subset, which is within statistical noise). GPT-5.5 and Claude Mythos sit even higher. If you need the absolute best performance for hard problems, you'll still reach for the proprietary models.

But most coding isn't hard problems. Most coding is moderately complex CRUD, wiring up APIs, fixing a misbehaving component, generating boilerplate. Devstral 2 handles that class of work perfectly well, and it does it without phoning home.

The takeaway

Mistral built something that shifts the ground under the coding agent market. An open-weight model that's competitive with proprietary alternatives, runs on a laptop you might already own, and costs nothing to use indefinitely. That was science fiction two years ago.

For no-code builders in particular, the local-first angle is the one that matters. If you've been sitting on the sidelines of AI-assisted development because your client data can't leave your machine, or because the API bills made you nervous, the wait is over. Grab Ollama, pull `devstral-small-2`, and try it.

It's not going to beat Claude on a SWE-bench leaderboard. But it will build your app, on your hardware, without telling anyone what you're building. In a lot of cases, that's exactly what you need.