Mistral AI

Z.ai's GLM-5.3 (August 2026), GLM-5.2's base with heavier post-training for coding and agentic work and a 1M-token context, served by Alibaba Cloud, Mistral and Together.

Rolling OCR 4 alias, pointing to OCR 4.1 since July 2026: Mistral's current document model, which adds paragraph-level bounding boxes, structural block labels and block-level confidence scores to Markdown extraction.

OCR 4.1, released in July 2026 and generally available since 2026-08-31, is Mistral's newest OCR model; it extracts document text with paragraph-level bounding boxes, block labels and per-block confidence scores, at $4 per 1,000 pages.

Moving alias for Mistral's newest OCR model, OCR 4.1 as of October 2026; it switched from OCR 4.0 in July 2026, so pin mistral-ocr-4-1 if output-format stability matters.

Z.ai's June 2026 open-weight (MIT) model for long-horizon coding, the first GLM to deliver a usable 1M-token context, with 128K output and selectable reasoning effort. Superseded as flagship by GLM-5.3, a post-trained update of the same base.

Mistral Medium 3.5 (April 2026), a dense 128B open-weight model under a modified MIT license that merges chat, reasoning and coding with a 256K context; it replaced Medium 3.1, Magistral and Devstral 2.

Voxtral TTS (March 2026), Mistral's first text-to-speech model: about 4B parameters, nine languages, preset voices and zero-shot cloning from a few seconds of reference audio, with roughly 70 ms model latency; open weights are CC BY-NC 4.0.

Latest alias for Mistral's Voxtral TTS, currently the 26.03 release; it speaks English, French, German, Spanish, Dutch, Portuguese, Italian, Hindi and Arabic and can clone a voice from a short sample without a transcript.

Mistral Small 4 (March 2026), an Apache 2.0 MoE with 119B total and about 6.5B active parameters that merges instruct, reasoning (formerly Magistral) and coding (formerly Devstral), with a 256K context and per-request reasoning effort.

Mistral Moderation 2 (March 2026), Mistral's API classifier that scores text or conversations across 11 categories, from hate and self-harm to PII, adds jailbreak detection and reads up to 128K tokens; it replaced the 2024 moderation model.

Mistral's rolling alias for its dedicated transcription model, currently Voxtral Mini Transcribe 2 (February 2026): 13 languages, speaker diarization, word timestamps, context biasing for up to 100 terms and files up to three hours.

Pinned Voxtral Mini Transcribe 2 (voxtral-mini-2602, February 2026), the batch speech-to-text model that replaced July 2025's Voxtral Mini Transcribe, priced at $0.003 per minute with timestamps and keyword biasing.

Voxtral Mini Transcribe Realtime (February 2026), a 4B streaming speech-to-text model with Apache 2.0 open weights, 13 languages and a configurable delay from 240 ms to 2.4 s; at 480 ms Mistral puts it within 1-2% WER of its batch model.

Mistral OCR 3 (December 2025), returning Markdown with HTML tables that keep merged cells, and handling handwriting, forms and low-quality scans better than OCR 2. OCR 4.1 is now newer; OCR 3 remains for existing integrations.

Version alias for Mistral OCR 3, the December 2025 document model that added header/footer extraction, hyperlinks and table-format options; Mistral reports a 74% win rate over OCR 2, at half OCR 4.1's per-page price.

Explicit OCR 3.0 identifier for the mistral-ocr-2512 snapshot; it lacks the paragraph-level bounding boxes and confidence scores introduced with OCR 4 and 4.1, but stays supported at $2 per 1,000 pages.

Ministral 3 3B (December 2025), the smallest Ministral 3 model: an Apache 2.0 model with a 256K context, sized for phones, laptops and other edge hardware.

The mid-sized Ministral 3 model (December 2025): 8B parameters, Apache 2.0 and a 256K context, replacing the 2024 Ministral 8B and Mistral Nemo 12B as Mistral's efficient local option.

Ministral 3 14B (December 2025), the largest of Mistral's Apache 2.0 edge models, with a 256K context and performance Mistral compares to Mistral Small 3.2 24B.

Mistral Large 3 (December 2025), an Apache 2.0 open-weight MoE with 675B total and 41B active parameters and a 256K context. Mistral Large 4 (about 1T parameters, 1M context) entered public preview on 2026-10-06.