Mistral AI built its reputation on punching above its weight class — releasing smaller models that beat larger ones on benchmarks. By 2026, the lineup spans from a lean 7B model you can run on a laptop to frontier-class API models that compete with GPT-4o at lower cost. Here is how to put them to work.
What changed in 2026
- Mistral Large 2 became the top API tier. Comparable to GPT-4o on most benchmarks, with European data residency as a compliance differentiator for EU businesses.
- Codestral Mamba shipped. A Mamba-architecture code model with significantly faster inference for fill-in-the-middle tasks — used in IDE integrations like Continue.dev and VS Code plugins.
- Mistral OCR launched. A document parsing model that extracts structured data from PDFs with strong accuracy — filling a niche no prior Mistral model covered.
- La Plateforme expanded. Mistral's API platform added fine-tuning endpoints, batch inference, and function calling on all major models.
The Mistral model lineup
| Model |
Context |
Best for |
| Mistral 7B Instruct |
32K |
Fast local inference, simple tasks |
| Mixtral 8x7B Instruct |
32K |
Balanced quality + speed, local or API |
| Mistral Small (API) |
32K |
Cost-effective API tasks |
| Mistral Medium (API) |
32K |
Mid-tier quality at ~50% GPT-4o cost |
| Mistral Large 2 (API) |
128K |
Frontier quality, EU data residency |
| Codestral |
32K |
Code generation and fill-in-middle |
| Mistral Embed |
— |
Embeddings for RAG pipelines |
How to access Mistral
Option 1 — La Plateforme API
- Sign up at
console.mistral.ai.
- Generate an API key.
- Use the OpenAI-compatible endpoint:
from mistralai import Mistral
client = Mistral(api_key="YOUR_KEY")
response = client.chat.complete(
model="mistral-medium-latest",
messages=[{"role": "user", "content": "Your prompt here"}]
)
print(response.choices[0].message.content)
Or use the OpenAI SDK with base_url="https://api.mistral.ai/v1" — most OpenAI code works with one line changed.
Option 2 — Local via Ollama
ollama pull mistral # Mistral 7B
ollama pull mixtral # Mixtral 8x7B (~26 GB)
ollama run mistral
Mistral 7B runs on 8 GB VRAM (GPU) or 16 GB RAM (CPU, slower). Mixtral 8x7B needs ~24–28 GB RAM or VRAM.
Option 3 — Hugging Face
All weights are on mistralai/Mistral-7B-Instruct-v0.3 etc. Use with vllm, llama.cpp, or transformers.
How to start
- Pick the right model for your budget. For API work: Mistral Small for cost, Mistral Large 2 for quality. For local: Mistral 7B on laptops, Mixtral 8x7B on workstations.
- Use the system prompt. Mistral models are instruction-tuned and follow system prompts well — define role, format, and constraints there.
- For code tasks, use Codestral. Do not use the general model for fill-in-middle code completion; Codestral is trained specifically for it.
- Test function calling on Medium or Large. Function calling on the 7B model is usable but less reliable; upgrade tiers if reliability matters.
Pricing reference
| Model |
Input (per 1M tokens) |
Output (per 1M tokens) |
| Mistral Small |
~$0.20 |
~$0.60 |
| Mistral Medium |
~$2.70 |
~$8.10 |
| Mistral Large 2 |
~$3.00 |
~$9.00 |
| Codestral |
~$0.30 |
~$0.90 |
Prices as of mid-2026; check docs.mistral.ai/deployment/pricing for current rates.
Common mistakes
Using the general model for code completion. Codestral has a fill-in-the-middle token format the general models do not have; use the right tool.
Ignoring EU data residency. If your use case is EU-based and data sovereignty matters, Mistral is the only frontier-class provider with native EU data residency. Do not overlook this as a compliance feature.
Running Mixtral 8x7B on a laptop. It needs 24–28 GB of memory to load. Mistral 7B is the laptop model; Mixtral is for workstations or API.
Not using the OpenAI-compatible API. If you have OpenAI SDK code, switching base_url and model name is usually all you need. No reason to rewrite.
What to skip
- Mixtral 8x22B locally — at ~90 GB, it needs a multi-GPU workstation. Use the API or drop to Mixtral 8x7B for local runs.
- Mistral 7B for multi-step agentic tasks — the 7B model is weak on complex tool-use; use Medium or Large for anything with function calls in a loop.
- Mistral Embed for non-Mistral pipelines — it performs well but ensure you embed queries and documents with the same model to avoid vector space mismatch.
FAQ
Is Mistral fully open-source?
The 7B and Mixtral model weights are released under Apache 2.0. Mistral Large is closed and API-only. Check licensing per model.
Does Mistral support JSON mode?
Yes, on all major API models — set response_format={"type": "json_object"} in the API call.
How does Mistral compare to Llama 3?
Comparable at similar parameter counts; Mistral has stronger multilingual and European language performance; Llama 3 has a larger community and more tooling. Both are worth evaluating for your use case.
Can I fine-tune via the API?
Yes — La Plateforme supports fine-tuning on Mistral Small and Medium via the dashboard or API. You upload a JSONL dataset in OpenAI fine-tuning format.
Where to go next
See How to use Llama in 2026, How to use Qwen in 2026, and ChatGPT vs Claude vs Gemini vs Grok in 2026.