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

Devstral 2

123B dense open coding model, SOTA for code agents

Visit official page →

About

Mistral's next-generation agentic coding model family, released with the open-source Mistral Vibe terminal CLI. Devstral 2 is a 123B-parameter dense transformer with a 256K context window; the sibling Devstral Small 2 (24B, Apache 2.0) runs on consumer hardware (single RTX 4090 / 32GB Mac). Built for multi-file reasoning, dependency-aware code changes, and repository-scale agent workflows; fine-tunable and on-prem deployable. HF repo for the 123B: mistralai/Devstral-2-123B-Instruct-2512. Succeeds the Devstral Small 1.x line (24B, Apache 2.0, 46.8%→68% class performance). Note: Devstral 2's license is a modified MIT — check the actual license text before commercial use (see red flags).

Specs

Context window

256,000 tokens (native)

Pricing

Currently free via the Mistral API (provider-announced; may change). Announced post-free pricing: $0.40/$2.00 per 1M input/output tokens for Devstral 2 (123B); Devstral Small 2 (24B) at $0.10/$0.30 — per third-party reporting, not independently verified against Mistral's pricing page.

Free quota

Free via the Mistral API at launch; weights free to download from Hugging Face and Ollama.

API access

Yes — Mistral La Plateforme API (currently free), plus self-hosting via Hugging Face weights and Ollama.

Weights

Open weights

Released

December 9, 2025

Benchmarks

Provider-reported (Mistral, Dec 2025): 72.2% on SWE-bench Verified (Devstral 2, 123B); 68.0% (Devstral Small 2, 24B); claimed 'up to 7x more cost-efficient than Claude Sonnet at real-world tasks'. Mistral describes Devstral 2 as ~5x smaller than DeepSeek V3.2 and ~8x smaller than Kimi K2 while matching their SWE-bench scores. Independent: third-party reviews broadly corroborate the 123B fitting on a single H100-class GPU and the Small 2 on an RTX 4090, but no independently-run SWE-bench Verified replication was found at research time.

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

Use cases

codeautomationdevopstesting

Trust score — 8/10

/10 trustVerified

Scored 8/10 (verified tier, metrics-v1). Evidence: HF top match: mistralai/Devstral-2-123B-Instruct-2512; HF downloads 5,790 (peer-normalized); released 308 days ago. Manual review: announced paid pricing third-party-reported; no independent benchmark replications found. License: modified MIT (disputed revenue cap — check license text).

How the trust score is built

Full methodology →
Adoption · stars/downloads, log-scaled vs peers
7.3/10
Maintenance · recency, releases, issue load
7.0/10
Openness · license + docs
7.2/10
Manual review · sanity read + red-flag check
9.0/10
  • · HF top match: mistralai/Devstral-2-123B-Instruct-2512
  • · HF downloads 5,790 (peer-normalized)
  • · released 308 days ago
  • · manual review: announced paid pricing third-party-reported; no independent benchmark replications found

More in the devstral family

Sources