mistral

Package

@ai-sdk/mistral

Factories

AIProviders.mistral

Capabilities

• Language
• Embedding
• Transcription
• Speech

Conversations API

let mistral = try AIProviders.mistral()
let model = try mistral.conversation("mistral-small-latest")
let result = try await model.generate(.init(
    messages: [.user("Find recent astronomy news.")],
    tools: ["web_search": MistralTools.webSearch()]
))
print(result.text)

conversation(...) uses /conversations with typed Mistral options, system instructions, file/image input, function and provider tools, sources, usage and streaming event conversion. MistralTools.webSearchPremium() exposes premium search. Every call is stateless (store=false) and replays the full prompt; conversation or agent continuation IDs are not exposed. Existing languageModel(...) and chatModel(...) retain Chat Completions.

A schema plus application tools on Chat Completions uses the synthetic JSON tool; its arguments become output text without hiding real application calls. Generic embedding dimensions are forwarded unless provider-specific dimensions override them.

Reasoning configuration for future models

Unknown chat model IDs inherit support for top-level reasoning and the mistral.reasoningEffort override. The override keeps precedence. Known non-reasoning Codestral, Ministral, legacy Mistral Large and Voxtral IDs retain their previous unsupported behavior.

Voxtral 3 speech and transcription

Speech IDs such as voxtral-mini-tts-3-0 and voxtral-mini-tts-3 accept providerOptions.mistral.expressiveness from 1 through 5 and phonemeHints, a dictionary mapping words to nonempty arrays of IPA phonemes. Inline delivery tags remain in the input text. Existing refAudio still takes precedence over voice.

let speech = try mistral.speech("voxtral-mini-tts-3")
let audio = try await speech.speak(.init(
    text: "[emotion=happy] I like tomato soup.",
    voice: "gb_jane_neutral",
    providerOptions: ["mistral": [
        "expressiveness": 1.5,
        "phonemeHints": ["tomato": ["t", "ə", "m", "ɑ", "t", "oʊ"]]
    ]]
))

Voxtral 3 transcription supports voxtral-mini-transcribe-3 and voxtral-small-transcribe-3, including their -3-0 and -latest aliases. returnLanguage: true requests detected language and cannot be combined with a pinned language. Context bias accepts nonempty phrases and punctuation, including more than 100 entries. Language plus timestamps is rejected only for the known legacy voxtral-mini-2507, voxtral-mini-2602 and voxtral-mini-latest IDs. Current, future and custom model IDs forward that combination. The API validates model-specific word timestamps and diarization requirements.

Notes

Mistral chat, embeddings, Voxtral transcription, and speech mirror upstream Mistral 4.0.72: MISTRAL_API_KEY, custom base URL and headers, versioned user-agent suffix, callable/language/chat, embedding/textEmbedding, transcription, and speech/speechModel aliases, chat /v1/chat/completions, embeddings /v1/embeddings, transcription /v1/audio/transcriptions, speech /v1/audio/speech, structured outputs, JSON-object instruction injection, stop sequences via native stop, top-level reasoning mapped to supported Mistral reasoning effort with providerOptions precedence, provider option schema validation including speech refAudio, tool/tool-choice mapping, fragmented streaming tool-call buffering, tool-result output conversion, image/PDF user file support, typed assistant thinking blocks, Voxtral option validation and rich segment/usage metadata, streaming text/reasoning/tool lifecycle, promptCacheKey to prompt_cache_key mapping, complete raw usage and cached-token accounting, speech reference-audio metadata redaction, response metadata, Mistral error schema, and unsupported image family. ProviderV4 type names, ESM-only packaging, Node 22 engines, optional JS headers config, and workflow serialization helpers are JS-only upstream concerns. The 2026-10-04 pass maps error finish reasons, injects JSON schemas when structuredOutputs is false, and retains native transcription usage. The 2026-10-11 pass: conversation(modelID) provides stateless /conversations generation and SSE with replayable builtin tools, sources, usage and errors. MistralTools.webSearch/webSearchPremium, Large 4 reasoning and common embedding dimensions are available. The 2026-10-18 pass: Known non-reasoning exclusions retain warnings while future model IDs inherit reasoning-effort support. The 2026-10-25 pass: Voxtral 3 speech forwards validated expressiveness and IPA phoneme hints; transcription supports detected language, phrase context bias and model-aware language/timestamp validation. Missing chat completion signals are interrupted.

Factory argument requirements are defined by the public Swift factory signatures. Use Public symbols when you need the exact initializer or factory declaration.

Return to the provider matrix.