Mistral AI news & updates

Mistral releases four Timestral forecasting models

Mistral released Timestral in four sizes — 5M, 28M, 123M and 209M parameters — for probabilistic multivariate forecasting from past-only and known-future covariates, with context windows up to 32,768 steps and 21 output quantiles. The repository code is Apache 2.0; the weights use Mistral’s Modified MIT terms, which do not authorise weight use where the user’s company or employer exceeded $20 million in global consolidated monthly revenue in the preceding month. Mistral directs those users to a commercial licence or hosted service. Mistral reports that the 209M checkpoint leads other pretrained zero-shot models on GIFT-Eval, TIME and fev-bench, but says a fuller technical report is still forthcoming.

Why it matters: A compact, downloadable model family gives teams a new general-purpose option for forecasting demand, energy, finance and other ordered data, although larger commercial users need to check the weight licence.

Official article dated 15 October 2026 without a publication time. The public repository’s initial release commit is timestamped 14:59:55 UTC and Hugging Face files were finalised later that afternoon; benchmark rankings are Mistral-reported rather than an independent audit.

Mistral brings Arabic realtime transcription to a WebGPU browser demo

Mistral’s official Hugging Face Space gained a working Arabic realtime transcription demo on 12 October, with the first app commit timestamped 09:59 BST. Its landing page describes local processing through Transformers.js and WebGPU, about 2.8 GB of quantized ONNX downloads, and a browser requirement for shader-f16 support. The demo uses an ONNX conversion linked from Mistral’s Voxtral Mini 4B Realtime Arabic model card. This is a new way to access an existing Arabic checkpoint, whose downloadable weights use Apache-2.0, rather than a newly released model or hosted API.

Why it matters: Readers can explore streaming Arabic transcription on compatible hardware while keeping the browser demo, downloadable weights and hosted-service availability distinct.

Mistral Vibe 2.26.1 adds deadline controls and resumable headless runs

Mistral Vibe 2.26.1 adds ways to supply headless prompts from a file or standard input and stop a run at a configured deadline. With an output directory set, each exit saves an export and session journal; runs interrupted by a deadline, SIGTERM or Ctrl-C can be resumed. Controls can disable background processes and subagents, and the new Document Library connector is experimental and off by default. The release also removes the bundled Devstral local model.

Why it matters: Automation becomes easier to bound and recover. Users relying on the removed local-model option need to review their configuration rather than assume it remains available.

Mistral Large 4 enters public preview

Mistral opened public API preview access to Large 4, a multimodal model with a one-million-token context window. The company says it targets coding, agent workflows and professional document work. Downloadable weights were not yet available when checked on 8 October; Mistral plans to release them at the end of October.

Why it matters: The preview gives developers access to Mistral’s new flagship ahead of its planned downloadable weights release.

Mistral retires OCR 4.0 and points users to OCR 4.1

Mistral’s 29 September changelog records the retirement of OCR 4.0 on 30 September and directs developers to OCR 4.1 at the same price. OCR 4.1 had already become generally available on 31 August, so this update is a migration notice rather than a new model launch.

Why it matters: Developers that selected OCR 4.0 explicitly should check their document processing integrations.