Second-generación abierto ingeniería de software agente con 262K contexto, profundo terminal agente capacidades, y hardened patch generación.
Technical Specifications
- Provider
- Mistral AI
- Category
- flagship
- Release Date
- 2025-12
- Batch API Discount
- 50% off async batch
- Consumer Subscription
- Free / Self-Hosted
- Reasoning Effort Control
- Configurable (Low / Med / High)
- Structured Outputs
- Yes
- Output Modalities
- text
- Official Documentation
- Visit Official Docs ↗
Data Privacy, Licensing & Deployment Governance
Official Provider Release Intelligence, API Quotas & Agentic Specs
- Official API Model ID (SDK String)
- devstral-2-latest
- Supported SDK & API Protocol
- OpenAI-Compatible REST / vLLM / SGLang / HuggingFace TGI
- Lifecycle & Deprecation Status
- Active GA (Guaranteed support through 2027+)
- Production Rate Limits (RPM & TPM)
- Unlimited (Hardware-bound on self-hosted GPU) / 10,000 RPM Cloud · Unlimited local throughput / 10M+ TPM Managed Cloud
- Long-Context Pricing Tier (>128K / >200K)
- Flat rate ($1.5/1M in) across full 262K window
- Prompt Cache Write & TTL Policy
- Cached input at $0.35/1M tokens
- Computer Use & Browser GUI Agent
- API / CLI Tool-Use & DOM Agent ready (no raw pixel desktop click loop)
- MCP & Parallel Tool Calling
- Native MCP Client/Server ready · Parallel function calling · 100% Strict JSON Schema
- Web Search Grounding & Code Sandbox
- Supported via external Tavily/Exa Search & E2B/Docker Sandbox tools
- Fine-Tuning & Distillation Support
- Full Weights SFT, LoRA, QLoRA, DPO & GRPO RL (Unsloth / Axolotl / Llama-Factory)
- Multimodal Tokenization Billing
- Text BPE (~4 chars/token)
- Weights Precision & Quantization
- BF16 / FP16, FP8 Dynamic, INT4 AWQ / GPTQ, GGUF (Q4_K_M / Q8_0)
- Minimum GPU VRAM Requirement
- 76 GB VRAM (4-bit Quantized) · 264 GB VRAM (Native BF16)
- Supported Inference Engines
- vLLM, SGLang, Ollama, llama.cpp, TensorRT-LLM, HuggingFace TGI
- System Card & Safety Framework
- Open-Weights System Card + Customizable Guardrail Layer (Llama Guard / ShieldGemma)
- Enterprise Compliance (SOC2 / HIPAA / ZDR)
- 100% Air-Gapped On-Premises / Private VPC · Full HIPAA, GDPR & SOC2 Control
Devstral 2 Cost, Context, Hardware & API Integration
For single-user chat, a $20/mo subscription caps your bill below this API volume.
Break-even point: ~81 requests/day equals $20/month.
Same Monthly Workload Cost Across Popular Reference Models
Live Prompt Paste Tokenizer & Cost Counter
Paste real system instructions, JSON schemas, or code to measure BPE token counts across tokenizers.
Hybrid Prompt Routing Cascade Simulator
Route routine queries to a fast lightweight model ($0.15/1M) and escalate only complex reasoning turns to Devstral 2:
Benchmark Scores & 6-Axis Capability Radar
- SWE-bench Verified
- 58.2%
- GPQA Diamond
- 66.5%
- MMLU-Pro
- 83.5%
- HumanEval
- 94%
- Terminal-Bench
- 52%
- AIME
- 78%
- Long-Context Reasoning
- 82%
- Chatbot Arena Elo
- 1365pts
Strengths
- Expanded 262K ventana de contexto un ingest entire multi-package monorepos
- de vanguardia 58.2% SWE-bench score para un de pesos abiertos modelo
- Hardened against regression bucles y terminal syntax hallucination
Limitations & Trade-offs
- sustancial 123B parámetro memoria requirement para local ejecución
- API token costos son moderately alto para intensive multiturno bucles
Editorial Assessment
un líder de pesos abiertos ingeniería de software agente pushing el boundaries de autónomo codificación.
Best for: Whole-repositorio migrations, autónomo desarrollador agentes, y complejo arquitectura refactors
Compare Devstral 2
Frequently Asked Questions about Devstral 2
How much does the Devstral 2 API cost per 1M tokens?+
Devstral 2 by Mistral AI is priced at $1.5 per 1M input tokens and $4.5 per 1M output tokens. With prompt caching enabled, cached input tokens cost $0.35 per 1M tokens, and asynchronous Batch API requests receive a 50% discount.
Should I use a $20/month subscription or the Devstral 2 pay-as-you-go API?+
A flat $20/month consumer subscription is predictable for daily interactive chat and manual coding. However, if you are building software, running automated workflows, or sending low-to-moderate volume with prompt caching enabled, the pay-as-you-go API is typically far more cost-effective and offers higher rate limits without manual UI caps.
How many words and pages fit in Devstral 2's context window?+
Devstral 2 features a 262K tokens context window, which equals approximately 196,608 English words or ~524 standard single-spaced pages in a single prompt. Its maximum output generation per turn is 65,536 tokens.
Can I run Devstral 2 locally on my own GPU (Open Weights)?+
Yes, Devstral 2 (123B) releases open weights that can be self-hosted locally or on private cloud GPUs using vLLM, Ollama, SGLang, or llama.cpp.
Does Mistral AI train on data sent to the Devstral 2 API, and is commercial use allowed?+
100% Isolated when self-hosted on private GPUs or VPC. License tier: Permissive Open-Weights (Apache 2.0 / MIT Commercial). Full commercial ownership of generated outputs. Supported deployment channels include: Self-Hosted GPU (vLLM / Ollama), Private VPC, or Managed Serverless API.
How does Devstral 2 perform on coding and reasoning benchmarks?+
Devstral 2 achieves a composite Intelligence Index of 56/100, scoring 58.2% on SWE-bench Verified, 66.5% on GPQA Diamond, 83.5% on MMLU-Pro, and 1365 pts on Chatbot Arena Elo.
What is Devstral 2 best used for?+
Devstral 2 is best suited for: Whole-repositorio migrations, autónomo desarrollador agentes, y complejo arquitectura refactors. Assessment: un líder de pesos abiertos ingeniería de software agente pushing el boundaries de autónomo codificación.
Mistral AI Family Lineage, Routing Tiers & Upgrade Path
Step down to a lighter sibling for high-volume routing or step up to Mistral AI's flagship reasoning tier:
Mistral Medium 3.5
Mistral Small 4
Devstral Medium
Devstral Small
Mistral Small 3.2
Mistral Medium 3
AI Glossary & Architecture Concepts
View full glossary →Key machine learning concepts and benchmarks relevant to Devstral 2:
La unidad básica de texto que procesa un modelo de lenguaje, que normalmente equivale a 3 o 4 caracteres o aproximadamente 0,75 palabras. Por ejemplo, una frase típica se divide en varios tokens.
Un conjunto estandarizado de pruebas para evaluar y comparar objetivamente las capacidades de distintos modelos de IA en tareas como programación (SWE-bench) o matemáticas avanzadas (AIME).
La cantidad máxima de tokens que un modelo puede considerar a la vez, incluyendo tanto la entrada que lee como la salida que genera. Una ventana de 1.000.000 de tokens puede procesar unas 750.000 palabras en una sola petición.
Un modelo de IA entrenado para invertir cómputo adicional en pensar y validar hipótesis lógicas antes de responder, diseñado para problemas complejos de ingeniería y matemáticas.
Una técnica de razonamiento paso a paso donde el modelo genera pensamientos intermedios antes de formular la respuesta final, mejorando radicalmente la resolución de problemas matemáticos y lógicos.
La capacidad de un modelo de detectar la necesidad de una función externa y emitir argumentos JSON estructurados para ejecutar API, búsquedas web o consultas a bases de datos.
Las puntuaciones de benchmarks, precios por token y límites de tasa evolucionan rápidamente. Verifique siempre las especificaciones actuales en la documentación oficial del proveedor.