Mistral's Devstral 2, released December 9 2025, scores 72.2% on SWE-bench Verified - a benchmark that measures whether a model can close real GitHub issues, not just write tidy functions on command. That puts it within 5 points of Claude Sonnet 4.5 (77.2%), at roughly one-seventh the API cost. The part worth sitting with: the smaller sibling, Devstral Small 2 at 24 billion parameters, lands at 68.0% on the same benchmark - only four points behind the 123B flagship, and it runs on a single RTX 4090 or a 32 GB Mac.
Most coverage leads with the headline number and moves on. The decisions that actually matter - which size to pick, what the license says, how it fits inside an agent harness - get a paragraph at best. This post covers those.
What makes Devstral 2 a coding agent model, not a coding model
Devstral 2 is not trying to be a generic chatbot that happens to write code. Mistral positioned it as a software-engineering model built for code agents: exploring codebases, editing across files, using tools, and staying productive on longer, messier tasks than one-shot code generation.
That framing matters because it changes what you test. A typical code-completion benchmark rewards models that write correct functions from a docstring. Agent benchmarks like SWE-bench Verified reward something harder: read a real GitHub issue, navigate an unfamiliar codebase, write a fix that passes the existing test suite, and submit it. Devstral 2 is assessed primarily on SWE-bench Verified and SWE-bench Multilingual - the two benchmarks most directly tied to autonomous software engineering - using the OpenHands scaffold from All Hands AI.
Devstral 2 excels at the multi-step, tool-using workflows that define modern coding agents. The tool-calling success rate matches the best closed models, which means fewer failed attempts and cleaner execution. The 256K context window enables architecture-level reasoning across large codebases. Devstral understands framework dependencies, tracks state across multiple files, and can retry with corrections when something fails.
The 256K context window means it can load an entire mid-size repository into a single inference pass, which matters for tasks like cross-file refactoring or understanding a codebase's internal dependency graph.
Devstral 2 integrates with the major agentic coding scaffolds: OpenHands, SWE-agent, Cline, Kilo Code, and Claude Code as the model backend. Mistral also ships Mistral Vibe CLI alongside it, an open-source command-line tool supporting end-to-end automated programming.
The 24B vs 123B choice is closer than the size difference suggests
The flagship Devstral 2 is a 123B dense transformer - not a mixture-of-experts model, which means every parameter is active on every token. It is intended for high-end clusters and serious production use. It typically requires at least four H100-class GPUs for comfortable inference at FP8 precision, where the memory footprint is still around 130 GB. This is not consumer-grade hardware, but it is a realistic target for enterprise on-prem deployments and specialized clouds.
The Small 2 is a different calculation entirely. Devstral Small 2 is clearly sized for local use. With modern 8-bit and 4-bit quantization, it fits within the VRAM envelope of a single RTX 4090 or similar high-end consumer GPU, and can be pushed onto a 32 GB Mac in a pinch. In practice, that means SWE-bench-competitive performance is available in a form factor that can live on an engineer's workstation rather than a datacenter rack.
The benchmark gap between them is genuinely small for the parameter delta. The 68.0% scored by Devstral Small 2 is a bigger surprise - a 24B model within 4 percentage points of the 123B version, running on consumer hardware.
Devstral 2 (123B) and Devstral Small 2 (24B) are 5x and 28x smaller than DeepSeek V3.2, and 8x and 41x smaller than Kimi K2 - proving that compact models can match or exceed the performance of much larger competitors.
Read the license before you download the weights
This is where teams get caught. The two models ship under different licenses, and the difference matters at scale.
Devstral Small 2, the 24-billion parameter variant, is covered under a standard Apache 2.0 license. That's a gold standard in open source: no revenue restrictions, no fine print, no need to check with legal. Enterprises can use it in production, embed it into products, and redistribute fine-tuned versions without asking for permission.
Devstral 2, the flagship 123B model, is released under what Mistral calls a "modified MIT license." That phrase sounds innocuous, but the modification introduces a critical limitation: any company making more than $20 million in monthly revenue cannot use the model at all - not even internally - without securing a separate commercial license from Mistral.
The restriction applies to the model and any derivatives, modifications, or combined works based on it, whether provided by Mistral AI or by a third party.
One developer commenting on VentureBeat's launch coverage put it plainly: "calling the Devstral 2 license 'modified MIT' is misleading at best. It's a proprietary license with MIT-like attribution requirements." The criticism is fair. If your legal team sees "MIT" and stops reading, they will miss the revenue cap.
The practical split looks like this:
| Devstral 2 (123B) | Devstral Small 2 (24B) | |
|---|---|---|
| License | Modified MIT | Apache 2.0 |
| Revenue cap | >$20M/month needs commercial deal | None |
| Hardware floor | ~4× H100s (130 GB FP8) | Single RTX 4090 / 32 GB Mac |
| SWE-bench Verified | 72.2% | 68.0% |
| API: input / output | $0.40 / $2.00 per 1M tokens | $0.10 / $0.30 per 1M tokens |
| Fine-tune and ship? | Check with legal first | Go ahead |
The 123B uses a modified MIT license that restricts use by companies with global consolidated monthly revenue above roughly $20M. Devstral Small 2 (24B) is plain Apache 2.0 with no such restriction. If you are a startup under that threshold, both models are effectively open. If you are above it, the 123B requires a commercial agreement; the 24B does not.
Beagle in action#eng-platform, 10:22amWiring Devstral 2 into an agent harness
Because Devstral 2 is designed for agent frameworks rather than chat, how you scaffold it matters as much as which size you pick.
Many teams are no longer choosing between "AI" and "no AI." They are choosing operating models: frontier hosted systems, cheaper open-weight stacks, or coding-specific models that can sit inside agent frameworks, terminals, CI workflows, and private infrastructure. Devstral 2 is explicitly built for that last category.
A few things worth knowing before you start:
- OpenHands is the reference harness. Mistral's own SWE-bench numbers use OpenHands as the scaffold. If you want results that match the published benchmark, start there.
- Tool-call behavior is tuned, not bolted on. They are tuned specifically for agentic coding: exploring a repo, editing multiple files, running tools, and retrying on failure, rather than just single-snippet completion. A simpler harness that doesn't give the model room to retry will underperform the benchmark numbers.
- The 24B fits in CI. A 4090-class node running Devstral Small 2 locally processes code without sending it to an external API. For teams whose security policy prevents source code leaving the network, that is the whole argument.
- Cline and Kilo Code ship native support. The launch is supported by partnerships with agent tools Kilo Code and Cline, so if your team is already on either, pointing at Devstral is a one-line config change.
- Terminal-Bench 2 scores are weaker. The Terminal Bench 2 score of 32.6% is worth flagging
- if your agent workflow is heavy on shell orchestration and general terminal tasks rather than repository-level code changes, the gap to frontier models is wider than SWE-bench suggests.
The non-obvious cost implication: at $0.40/$2.00 per million tokens input/output, Devstral 2 is approximately 7x cheaper than Claude Sonnet 4.5 at $3/$15. Run those numbers against an agent that touches 50K tokens per task across a 20-engineer team and the monthly bill difference is not marginal. For teams running agents at CI frequency - every commit, not just on demand - the self-hosted 24B is cheaper still, at zero marginal cost past hardware amortization.
Devstral 2 coding agent: common questions
What is Devstral 2 and how does it differ from a regular coding model?
Devstral 2 is not trying to be a generic chatbot that happens to write code. Mistral positioned it as a software-engineering model built for code agents: exploring codebases, editing across files, using tools, and staying productive on longer, messier tasks than one-shot code generation. The practical difference is that it is tuned to run multi-step loops inside agent scaffolds, not to return a single completion.
Which Devstral 2 size should I run locally?
The 123B is the pick when you have the hardware and need the extra few points on hard, multi-file refactors. For most self-hosters, start with the 24B and the free Cline + Ollama setup and only step up if the smaller model leaves capability on the table. The 4-point SWE-bench gap between them rarely decides real outcomes.
Can a large company freely use Devstral 2?
Devstral 2 (123B) is free for commercial use unless your company's global consolidated monthly revenue exceeds roughly $20 million, in which case Mistral's license says you are not authorized to use it without a commercial agreement. For small teams and individuals both models are usable commercially; only very large companies hit the 123B restriction. Devstral Small 2 (24B) carries no such cap.
How does Devstral 2 compare to Claude Sonnet on SWE-bench?
The 72.2% SWE-bench score puts Devstral 2 above DeepSeek V3.2 but below Claude Sonnet 4.5 at 77.2%. For a model that fits on two A100 80 GB GPUs and runs fully on-premises, that gap is narrow enough to matter. The cost difference - roughly 7x per token via API - makes the remaining benchmark gap workable for most engineering teams.
What agent scaffolds work with Devstral 2?
Devstral 2 integrates with OpenHands, SWE-agent, Cline, Kilo Code, and Claude Code as the model backend. Mistral also ships its own Mistral Vibe CLI for terminal-native use. For teams on OpenCode - the MIT-licensed harness that crossed 165K GitHub stars in 2026 - Devstral 2 works as one of the 75-plus supported model endpoints.