Skip to content
← Models
Open weightsMistral/10 trustVerified
Devstral 2 2512 logo

Devstral 2 2512

Open 123B agentic coding model with 262K context, aliased as devstral-latest on Mistral's API.

Visit official page →

About

Devstral 2 2512 is Mistral's open agentic coding model, released December 9, 2025. It is a 123-billion-parameter dense transformer with a 262,144-token context window, trained for exploring codebases and orchestrating changes across files. On the Mistral API it is aliased as devstral-latest and mistral-code-agent-latest, meaning it is the current recommended coding model, not a legacy checkpoint.

Third-party roundups from December 2025 put it at 72.2% on SWE-bench Verified, within reach of Claude Sonnet 4.5's 77.2%, and Mistral's human evaluations had it beating DeepSeek V3.2 in 42.8% of head-to-heads. Weights are open, and it runs through Hugging Face, vLLM, and Ollama (a 24B Small 2 variant fits a single RTX 4090). API pricing is $0.40 per million input and $2.00 per million output tokens.

Specs

Context window

262,144 tokens

Pricing

$0.40 per million input tokens, $2.00 per million output tokens; cache reads $0.04 per million

Free quota

Free tier via OpenRouter's :free endpoint

API access

Mistral API (devstral-2512, alias devstral-latest), Hugging Face, Ollama, vLLM

Weights

Open weights

Released

December 9, 2025

Benchmarks

  • SWE-bench Verified: 72.2% (Dec 2025, third-party roundup)
  • Human preference: won 42.8% of head-to-heads vs DeepSeek V3.2 (28.6% losses), per Mistral

Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.

Use cases

agentic-codingcodebase-explorationrefactoringself-hosted-coding

Trust score — 8/10

/10 trustVerified

Scored 8/10 (verified tier, metrics-v1). Evidence: aliased as devstral-latest and mistral-code-agent-latest on Mistral's live API with community distribution via Ollama and vLLM. Open weights. Release facts corroborated across OpenRouter, ModelIndex, and Mistral's own API; docked for Mistral-reported human preference figures.

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
8.0/10
Maintenance · recency, releases, issue load
8.0/10
Openness · license + docs
10.0/10
Manual review · sanity read + red-flag check
8.0/10
  • · Adoption 8: aliased devstral-latest/mistral-code-agent-latest on Mistral API; Ollama and vLLM community usage; free OpenRouter tier
  • · Maintenance 8: released Dec 9, 2025; active and not deprecated per Mistral's own API
  • · Openness 10: open weights (0.7*10) + public Hugging Face/Ollama distribution (0.3*10)
  • · Manual 8: release facts (Dec 9 2025; 123B dense; 262,144 context; $0.40/$2.00; SWE-bench Verified 72.2%) corroborated across OpenRouter, ModelIndex, and Mistral's live API; docked for Mistral-reported human preference figures

More in the devstral family

Sources