Apr 17, 2026 · 4 min read

Last updated on Apr 19, 2026

Devstral 2 Complete Guide — Mistral's Open-Source Coding Agent Model (2026)

📢 Update: Mistral Medium 3.5 has replaced Devstral 2 as the default model in Vibe CLI. See the Medium 3.5 complete guide and Vibe 2.0 remote agents guide.

Devstral 2 is Mistral AI’s dedicated coding agent model — a 123B dense transformer with a 256K context window that scores 72.2% on SWE-bench Verified. That puts it among the best open-weight coding models available, competing directly with Claude Code and GLM-5.1.

Unlike Codestral (optimized for autocomplete), Devstral 2 is built for agentic coding — planning, executing, debugging, and iterating across entire repositories.

Specs

Spec	Devstral 2	Devstral Small 2
Parameters	123B (dense)	24B (dense)
Context window	256K	256K
SWE-bench Verified	72.2%	~58%
Architecture	Dense transformer	Dense transformer
License	Modified MIT	Modified MIT
Quantized size	~65GB (Q4)	~14GB (Q4)

The 256K context window is the standout feature — it can reason across entire codebases in a single pass. Most competing models top out at 128-200K.

Devstral 2 vs the competition

Model	SWE-bench	Context	Params	License	Local?
Devstral 2	72.2%	256K	123B	Mod. MIT	1x H100
GLM-5.1	58.4 (Pro)	200K	754B MoE	MIT	4x A100
Claude Opus 4.6	72.1%	200K	Unknown	Proprietary	No
Kimi K2.5	65.8%	256K	1T MoE	MIT	4x A100
Mistral Large 2	~68%	128K	123B	Research	1x H100

Devstral 2 essentially matches Claude Opus on SWE-bench while being open-weight and runnable on a single server node. That’s remarkable.

How to use Devstral 2

Via Mistral API

from mistralai import Mistral

client = Mistral(api_key="your-key")
response = client.chat.complete(
    model="devstral-2-latest",
    messages=[{"role": "user", "content": "Refactor this module to use dependency injection"}]
)

Via OpenRouter

from openai import OpenAI
client = OpenAI(base_url="https://openrouter.ai/api/v1", api_key="your-key")
response = client.chat.completions.create(
    model="mistralai/devstral-2",
    messages=[{"role": "user", "content": "Fix the race condition in this handler"}]
)

See our OpenRouter guide for full setup.

With Aider

aider --model mistralai/devstral-2 --openai-api-base https://api.mistral.ai/v1

See our Aider guide for configuration details.

With Mistral Vibe CLI

Mistral released their own terminal coding tool alongside Devstral 2:

npm install -g @mistralai/vibe-cli
vibe

Vibe CLI is purpose-built for Devstral and includes project-aware context management.

Running locally

Devstral 2 at 123B is the same size as Mistral Large 2 — it fits on a single high-end GPU:

Setup	Speed	Usable?
1x H100 (80GB)	~30 tok/s	✅ Excellent
2x A100 (160GB)	~25 tok/s	✅ Good
Mac Studio Ultra 192GB	~5-8 tok/s (Q4)	⚠️ Slow but works

For consumer hardware, use Devstral Small 2 (24B) instead — it runs on a single RTX 4090 or a Mac with 32GB RAM:

ollama pull devstral-small:24b
aider --model ollama/devstral-small:24b

Devstral 2 vs Codestral — what’s the difference?

Both are from Mistral, but they serve different purposes:

	Devstral 2	Codestral
Purpose	Agentic coding (agent)	Autocomplete (FIM)
Size	123B	22B
Context	256K	256K
Best at	Multi-file refactors, bug fixes	Inline completions, tab suggestions
Use with	Aider, Vibe CLI	Continue.dev, IDE

The ideal Mistral setup: Devstral 2 for complex agent tasks + Codestral for fast autocomplete.

Bottom line

Devstral 2 is the best open-weight coding agent model in 2026. It matches Claude Opus on SWE-bench, has the largest context window (256K), and runs on a single server. The modified MIT license allows commercial use. If you’re building AI coding tools or need a self-hosted coding agent, Devstral 2 should be your first choice.

FAQ

Is Devstral 2 free?

Yes, Devstral 2 is released under a modified MIT license that allows free commercial use. You can download the weights and self-host without any licensing fees.

Can I run Devstral locally?

Yes, Devstral 2 is designed to run on a single server with high-end GPUs. It’s one of the most practical frontier coding models for local deployment, requiring approximately 2x A100 80GB GPUs for full precision.

How does Devstral compare to Codestral?

Devstral 2 is optimized for agentic coding workflows (multi-file edits, tool use, autonomous problem-solving), while Codestral focuses on fast code completion and generation. Devstral 2 significantly outperforms Codestral on SWE-bench and complex coding tasks.

Is Devstral good for coding?

Devstral 2 is specifically built for coding and matches Claude Opus on SWE-bench verified. It excels at autonomous code editing, debugging, and multi-step development tasks with its 256K context window.