by MistralReleased 16 March 2026
Mistral Small 4 folds three earlier Mistral strengths into one model: step-by-step thinking from Magistral, picture reading from Pixtral and coding helpers from Devstral. You send text and pictures, up to about 262 pages, and get text back, with tool use, exact-form answers and a setting for how hard it thinks.
Reads about 260 pages · writes up to 200 pages
Standard OpenAI fields, grouped by what they control. Anything not listed is accepted and ignored.
Shaping the reply
How long, how varied, where to stop.
max_tokenstemperaturetop_pstopseedfrequency_penaltypresence_penaltyReasoning
Let it think before answering — and see the thinking.
reasoningreasoning_effortinclude_reasoningStructured output
Get JSON you can parse instead of prose.
response_formatstructured_outputsTools
Let the model call your functions.
toolstool_choiceNo. Mistral has retired it. Other Mistral models are on sale.
Mistral Small 4 folds three earlier Mistral strengths into one model: step-by-step thinking from Magistral, picture reading from Pixtral and coding helpers from Devstral. You send text and pictures, up to about 262 pages, and get text back, with tool use, exact-form answers and a setting for how hard it thinks.
Mistral says it supports dozens of languages, naming French and Arabic among them. Darja is not named.
It reads text and images. Up to 262K tokens in one request, about 260 pages. It writes up to 157 000 words at once.
Yes, its weights are public. Open weights on Hugging Face
Mistral. Released 16 March 2026.
Mistral's flagship speech-to-text model that turns audio into transcripts with word timestamps and speaker labels, with context biasing for domain-specific vocabulary
The Voxtral Mini derivative that only transcribes, taking audio from meetings, voice notes and podcasts and returning the text through a standard transcription request
Mistral's compact speech-to-text model that turns audio into transcripts with word timestamps and speaker labels, and can be steered with context biasing to catch domain terms
Mistral AI's text embedding model from late 2023, whose 1024-dimension vectors capture how passages relate, for semantic search and RAG; Codestral Embed is the version for code
From what its maker and two benchmark sites publish, and where it stands on OpenDunes. Read 26 Oct 2026.
Thinks first
Works through the problem before it answers.
Uses your tools
Can call functions your code provides.
Fills your JSON
Answers in the JSON shape you give it.
Searches the web No.
Can look things up online while it answers.
Cheaper on repeats
A prompt it saw a moment ago costs less the second time.
Open weights
The model's weights are public.
Fix a bug and explain the change
Thinks first
Run an agent that calls your inventory API
Uses your tools
Read a chart screenshot and return the values as JSON
Images · Fills your JSON