Devstral 2 2512

Tool callingStructured output

Hugging Face Website

Mistral's coding-agent model for repository work, terminal tasks, and software fixes

Context: 262K
Released: 10mo ago
Parameters: 123B
Input / 1M tokens: $0.40
Output / 1M tokens: $2.00
Cached input / 1M tokens: $0.04

Sources:models.devOpenRouter

Benchmarks

DeepSWE is the primary coding benchmark where an exact model evaluation exists. Other sourced evaluations appear below; missing scores are not zero.

Not evaluated in DeepSWE 1.1 as of 2026-09-22. Results are only shown for exact model versions in the official dataset. Official leaderboard.

No benchmark results have been matched to this model yet.

Specifications

Developer: Mistral AI
Model family: Devstral
Type: Language model
Predecessor: —
Successor: —
Released: December 9, 2025
Model ID: mistralai/devstral-2512
Openness: Open weights
License: —
Parameters: 123B
Input modalities: text, file
Output modalities: text
Modality: text+file->text
Native context: 256K
Tokenizer: Mistral
Max output tokens: 210K
Knowledge cutoff: 2025-12
Hugging Face: mistralai/Devstral-2-123B-Instruct-2512
Source freshness: Metadata synced Oct 8, 2026, 10:17 PM UTC

Open-weight builds

Ollama variants

Every published Ollama tag matched to this model, including parameter editions, instruction/base variants, and quantized builds.

3 builds

Variant	Parameters	Type	Quantization	Size	Context	Input
devstral-2:123b-instruct-2512-fp16	123B	instruct	FP16	250 GB	256K	text
devstral-2:latest	123B	instruct	Q4_K_M	75 GB	256K	text
devstral-2:123b-instruct-2512-q8_0	123B	instruct	Q8_0	133 GB	256K	text

Additional source evidence

Source model: mistral/devstral-2512
Source family: devstral
Last updated: 2025-12-09
Knowledge cutoff: 2025-12
Model context: 262K
Model output limit: 262K
Open weights: Yes
File attachments: No
Reasoning: No
Tool calling: Yes
Temperature control: Yes

Downloadable weights

Hugging Face

Published benchmarks

Benchmark	Score	Evaluation details	Evidence
Artificial Analysis Coding Index	23.7 · index	2026-05-31	Source
SciCode	33.1 · percent correct	2026-05-31	Source
Terminal-Bench Hard	18.9 · success rate	2026-05-31	Source

Complete source record (JSON)

{
  "id": "mistral/devstral-2512",
  "name": "Devstral 2",
  "description": "Mistral's coding-agent model for repository work, terminal tasks, and software fixes",
  "family": "devstral",
  "last_updated": "2025-12-09",
  "knowledge": "2025-12",
  "attachment": false,
  "reasoning": false,
  "tool_call": true,
  "temperature": true,
  "modalities": {
    "input": [
      "text"
    ],
    "output": [
      "text"
    ]
  },
  "limit": {
    "context": 262144,
    "output": 262144
  },
  "benchmarks": [
    {
      "name": "Artificial Analysis Coding Index",
      "score": 23.7,
      "source": "https://openrouter.ai/mistralai/devstral-2512/benchmarks",
      "metric": "index",
      "date": "2026-05-31"
    },
    {
      "name": "SciCode",
      "score": 33.1,
      "source": "https://openrouter.ai/mistralai/devstral-2512/benchmarks",
      "metric": "percent correct",
      "date": "2026-05-31"
    },
    {
      "name": "Terminal-Bench Hard",
      "score": 18.9,
      "source": "https://openrouter.ai/mistralai/devstral-2512/benchmarks",
      "metric": "success rate",
      "date": "2026-05-31"
    }
  ],
  "release_date": "2025-12-09",
  "open_weights": true,
  "weights": [
    {
      "label": "Hugging Face",
      "url": "https://huggingface.co/mistralai/Devstral-2-123B-Instruct-2512"
    }
  ]
}

Infrastructure

Provider availability

All providers

Provider	Endpoint	Context	Input /M tokens	Output /M tokens	Cached input /M tokens	Cache write /M tokens	Uptime 1d	Data handling
Mistral	mistral/eu	262K	$0.40	$2.00	$0.04	—	100.00%	ZDR: Yes Training: Unknown Prompt logging: Unknown Retention: Unknown ZDR source Privacy policy Checked 2026-10-15

Provider data:OpenRouter

Deployment graph

Can I run this?

Vertical Tabs only shows compatibility when it has structured evidence tying this exact model to hardware, runtime, quantization, memory use, and observed performance.

Browse hardware