Mistral AI 推出的 Codestral 是一款面向本地私有化部署与代码生成的开源权重编程大模型，具备 256K 上下文窗口、110 Token/秒生成速度与 40/100 综合智能指数（每百万 Token $0.2 输入 / $0.6 输出）。
Technical Specifications
- Provider
- Mistral AI
- Category
- coding
- Release Date
- 2024-05
- Batch API Discount
- 50% off async batch
- Consumer Subscription
- Free / Self-Hosted
- Reasoning Effort Control
- Fixed
- 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)
- codestral-latest
- Supported SDK & API Protocol
- OpenAI-Compatible REST / vLLM / SGLang / HuggingFace TGI
- Lifecycle & Deprecation Status
- Legacy Stable (Upgrade path available in Mistral AI lineup)
- 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 ($0.2/1M in) across full 256K window
- Prompt Cache Write & TTL Policy
- Cached input at $0.05/1M 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
- 14 GB VRAM (4-bit Quantized) · 47 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
Codestral Cost, Context, Hardware & API Integration
Pay-as-you-go API is cheaper for this volume (saves $3.50/mo vs $20 flat plan).
Break-even point: ~606 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 Codestral:
Benchmark Scores & 6-Axis Capability Radar
- SWE-bench Verified
- 28%
- GPQA Diamond
- 42%
- MMLU-Pro
- 70%
- HumanEval
- 81.1%
- Chatbot Arena Elo
- 1240pts
Strengths
- 提供 256K 上下文窗口与最高 8K 输出容量，核心基准表现优异（SWE-bench Verified 28% · GPQA Diamond 42%）
- 高达 110 Token/秒的生成吞吐量，首字延迟 (TTFT) 仅约 350 毫秒
- 开放模型权重 (22B)，支持通过 vLLM、SGLang 或 Ollama 在私有 GPU 集群上本地部署与微调
Limitations & Trade-offs
- 在本地以满血精度运行需要较高的 GPU 显存 (VRAM) 与运维基础设施
- 在超长链路的前沿科研推导任务上，深度略逊于顶配重型旗舰推理模型
Editorial Assessment
Codestral 凭借 40/100 的智能得分、256K 上下文容量与 110 Token/秒的响应速度，在 Mistral AI 模型矩阵中提供了极佳的性能与成本平衡。
Best for: 面向本地私有化部署与代码生成的开源权重编程任务、企业级 RAG 检索增强生成、自动化智能体工作流与高可靠 API 集成。
Compare Codestral
Frequently Asked Questions about Codestral
How much does the Codestral API cost per 1M tokens?+
Codestral by Mistral AI is priced at $0.2 per 1M input tokens and $0.6 per 1M output tokens. With prompt caching enabled, cached input tokens cost $0.05 per 1M tokens, and asynchronous Batch API requests receive a 50% discount.
Should I use a $20/month subscription or the Codestral 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 Codestral's context window?+
Codestral features a 256K tokens context window, which equals approximately 192,000 English words or ~512 standard single-spaced pages in a single prompt. Its maximum output generation per turn is 8,192 tokens.
Can I run Codestral locally on my own GPU (Open Weights)?+
Yes, Codestral (22B) 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 Codestral 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 Codestral perform on coding and reasoning benchmarks?+
Codestral achieves a composite Intelligence Index of 40/100, scoring 28% on SWE-bench Verified, 42% on GPQA Diamond, 70% on MMLU-Pro, and 1240 pts on Chatbot Arena Elo.
What is Codestral best used for?+
Codestral is best suited for: 面向本地私有化部署与代码生成的开源权重编程任务、企业级 RAG 检索增强生成、自动化智能体工作流与高可靠 API 集成。. Assessment: Codestral 凭借 40/100 的智能得分、256K 上下文容量与 110 Token/秒的响应速度，在 Mistral AI 模型矩阵中提供了极佳的性能与成本平衡。
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 2
Devstral Medium
Devstral Small
Mistral Small 3.2
AI Glossary & Architecture Concepts
View full glossary →Key machine learning concepts and benchmarks relevant to Codestral:
The basic unit of text that a language model processes, typically representing about 3 to 4 English characters or roughly 0.75 words. For example, the sentence "Artificial intelligence is fast" breaks down into about 5 tokens.
A standardized evaluation suite used to objectively measure and rank an AI model's capabilities across specific tasks — such as SWE-bench Verified for software engineering, AIME for competitive math, or MMLU for general knowledge.
The maximum number of tokens a model can consider at once — both the input it reads and the output it writes. A 1,000,000-token context window can hold roughly 750,000 English words in a single request.
The ability of a language model to recognize when an external capability is needed and output structured JSON arguments to invoke external APIs, databases, or terminal commands. The runtime executes the tool and passes results back to the model.
An open standard created by Anthropic that allows AI applications and agents to connect seamlessly to local filesystems, development environments, and third-party APIs via standardized server interfaces.
A cost and latency optimization where inference providers store pre-computed attention states for repeated prompt prefixes. Reusing cached tokens typically slashes API input costs by 50% to 90% and cuts TTFT substantially.
Benchmark scores, token prices, and rate limits evolve rapidly as AI labs release model updates. Always verify current specifications and pricing on the official provider documentation.