Mistral AIPublic preview
Voice-agent SDK with transcription and speech models for live calls. Mini Transcribe 3 batch is listed at $0.0035 a minute and Speech 3 at $13 per million characters. The company says transcription covers more than 50 languages. Those benchmark rankings are Mistral's own.
Latest update Oct 21, 2026
01
API release on Oct. 20, 2026. Mistral's model docs label the speech models Public Preview. On-premises deployment is by request.
02
On Oct. 20, 2026 Mistral released Voxtral 3, a voice-agent SDK plus transcription and speech models, on its API and for self-deployment. Mini Transcribe 3 batch is listed at $0.0035 a minute, Small Transcribe 3 at $0.004 a minute, and Speech 3 at $13 per million characters. The company says transcription covers more than 50 languages; the benchmark rankings are Mistral's own.
03
Mistral's frontier model generation, headlined by Mistral Large 3 - a 675B-parameter sparse mixture-of-experts model (41B active) released open-weight under Apache 2.0 in December 2025 - alongside Mistral Medium 3.5 (the 'first flagship merged model,' default in Vibe since May 2026) and the efficient Mistral Small 4. A next-generation 'fat but sparse' MoE frontier model entered partner early access in July 2026.
Mistral's consumer and enterprise AI assistant, rebranded from Le Chat to Vibe on May 28, 2026 and repositioned as 'one agent for long-running, multi-step work.' Work Mode connects to Google Workspace, Outlook, SharePoint, Slack and GitHub to execute multi-step tasks autonomously; Code Mode runs persistent sandboxed coding sessions through to merged pull requests, with a VS Code extension and CLI.
Sovereign European AI cloud built with Nvidia: GPU capacity for training and inference operated on Mistral's own data centers, anchored by the 44 MW Bruyères-le-Châtel site (13,800 Nvidia GB300 GPUs) in Essonne, a 10 MW inference facility at Les Ulis due in Q3 2026, a €1.2B second site under construction with EcoDataCenter in Borlänge, Sweden (250-600 MW, Nvidia Vera Rubin GPUs, opening 2027) - Mistral's first infrastructure investment outside France - and the 'Campus AI' joint venture with MGX, Bpifrance and Nvidia - a 1.4 GW flagship site at Fouju now under construction, scaled in June 2026 to a roughly 3 GW nationwide network with a second site pending.
Mistral's developer and enterprise platform: pay-per-token API access to the full model catalog plus AI Studio for building production AI - Workflows (April 2026), built-in and custom MCP connectors, Search Toolkit, prompt/skills versioning and governance (July 2026), and Agentic Search (August 2026). Forge, launched March 2026, lets enterprises build frontier-grade custom models on Mistral's stack.
A broad lineup of purpose-built models: Codestral and Devstral for coding, Magistral for step-by-step reasoning, Voxtral for speech (with an open-weight TTS model added March 2026), Mistral OCR 4 for document intelligence (June 2026), Leanstral for formal proofs, Robostral Navigate (July 2026) - an 8B embodied-navigation model steering robots from a single RGB camera - and Shieldstral (August 2026), a 3B open-weight safety classifier that judges text and images against plain-language moderation policies at inference time instead of a fixed, retrained taxonomy.
Public-preview natively multimodal mixture-of-experts model. Mistral's 6 Oct 2026 blog calls it a 1-trillion-parameter model with 52 billion active parameters and says the preview API is on Mistral Studio, with open weights due by the end of October. The docs page lists API identifier mistral-large-4+1, 52B active parameters, 1.05T total parameters, a 1.6B vision encoder, and a 1M context window. The Hugging Face card, still an upcoming release on 16 Oct 2026, says 49 billion active parameters per token and 52 billion including embeddings and output layers, with a current ETA of 31 Oct 2026. Artificial Analysis describes the same preview as 1T total and 49B active, with a 512k context window, and scores it 38 on the Intelligence Index and 50 on the Cyber Index. Weights were not downloadable on 16 Oct 2026, so there is no open-weight release date or public license yet.
Time-series foundation models in 5M, 28M, 123M and 209M sizes. They output 21 forecast quantiles and accept past-only and known-future covariates. Mistral says the 209M model leads pretrained zero-shot results on GIFT-Eval, TIME and fev-bench.
04
Tesla
Tesla's highest-volume electric crossover SUV and the core of its global vehicle lineup. Tesla says it completed the refreshed Model Y rollout in 2025, and the current lineup spans rear-wheel drive, all-wheel drive, premium, and performance variants depending on market.
Tesla
Tesla's mass-market electric sedan and one of the company's longest-running products. The current Model 3 lineup emphasizes efficiency, software features, and lower entry pricing than Model Y or Cybertruck.
Tesla
Tesla's stainless steel electric pickup truck, positioned as a high-performance utility vehicle with adaptive suspension, high towing capacity, and fast charging. Tesla's current lineup includes Dual Motor All-Wheel Drive, Premium All-Wheel Drive, and Cyberbeast trims.
Tesla
Tesla's advanced driver-assistance software that can navigate city streets, make lane changes, park, and use summon features under active driver supervision. Tesla positions the same AI foundation behind FSD as part of its broader autonomy and robotics stack.