Mistral says Large 4 now has open weights and three reasoning modes. This guide helps you pick settings without overspending. Facts below come from Mistral’s release post; the prompts are ours.
Step 1: pick a reasoning mode
Mistral describes Low for simple tasks, Medium for everyday tasks and High (the default) for the hardest tasks. Higher modes usually cost more output tokens, so start low and move up only when answers fail.
- Low: classification, short rewrites, extraction.
- Medium: drafting, summaries, routine coding help.
- High: multi-step debugging, planning, tasks where a wrong answer is costly.
Step 2: pick hosted or self-hosted
- Hosted: Mistral lists its API, Mistral Studio, OpenRouter, Microsoft Foundry and NVIDIA-hosted endpoints. Mistral’s stated API price is $0.70 per million input and $2.10 per million output tokens. Confirm current prices on the host you choose.
- Self-hosted: weights on Hugging Face, run with vLLM or SGLang, or an NVIDIA NIM container. A 1 trillion parameter model needs serious GPU capacity, so price the hardware before you commit.
- Licence check: the licence is a modified MIT licence. If you sell model access as a service, Mistral says you need a separate commercial licence. See our open-weight checklist.
Step 3: run the same test at each mode
Use a task from your own work and compare accuracy and token use:
Here is a function and its failing test. Explain the cause in two sentences, then give a corrected version and one extra test. If you are unsure, say what you would check first.Summarise the attached notes in five bullets. Mark any bullet that is a guess rather than stated in the notes.Keep the cheapest mode that passes. Treat vendor benchmarks as a starting point: compare against other open-weight models and our DeepSeek vs Mistral comparison on your own data.