Mistral's Voxtral 3 Bundles Transcription, Expressive Speech and a LiveKit-Based Agent SDK Behind a Sovereignty Pitch -- the Same Day Microsoft Said Foundry Local Will Run Fully Disconnected

On October 20, 2026, Mistral launched Voxtral 3, "a new family of state-of-the-art speech-to-text and text-to-speech models, alongside a voice agent SDK". The post reached 539 points on Hacker News.

The models

• Transcription: Voxtral Small Transcribe 3 covers 50 languages, "including 24 official EU languages and many low-resource Asian and Indic languages". Its FLEURS evaluation set includes Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi and Telugu. Mistral says it outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe, and that it supports speaker diarization and context biasing, which feeds the model domain terms instead of fine-tuning it. The 4B-parameter Mini model costs $0.0035 a minute.
• Realtime: Voxtral Realtime Transcribe 3 promises "sub-500ms latency". The launch text says it covers "over 35 languages", and the product card says 40+.
• Speech: Voxtral Mini TTS 3 covers 35+ languages and 50+ dialects at $13 per million characters. Its voices adjust "emotion, pace, and emphasis based on conversational context." Pro Voice reproduces a speaker's style "using just 30 minutes of audio".

The agent SDK

The SDK is "built on LiveKit". Realtime transcription detects end of turn, a general model such as Mistral Small reasons and calls tools, and Voxtral speech replies. "No stage waits for the one before it, so each turn completes in under one second, and the agent stops talking when the caller cuts in." It handles phone calls over SIP and browser or app sessions over WebRTC. Because each stage is a separate model, the LLM, voice or language can be swapped without rewriting the agent. Mistral's audio samples show where voice agents usually fail: in one comparison, "21:00pm" is read out as "21,000pm" and "10/10" as "10 over 10".

Mistral sells all of this on control. It offers EU-hosted, US and global endpoints, plus self-deployment so organisations can decide "whether conversations are retained". On-premises deployment goes through its sales team. As of October 21, Mistral's Hugging Face organisation lists no Voxtral 3 weights.

Microsoft makes the same pitch

The same morning, Douglas Phillips, president of Microsoft Specialized Clouds, said Foundry Local will become generally available at Ignite "for connected and disconnected operations on Azure Local", with Microsoft, partner and open-weight models. Developers will be able to use "GitHub Copilot with locally hosted models for inference". External key management with Azure Managed HSM is now generally available, and Data Guardian adds EU-resident oversight of Microsoft engineers' remote access for EU and EFTA customers.

What remains uncertain. The speech-quality charts are labelled "Internal eval" and were judged by gemini-3.8-flash, one vendor's model grading others. No independent latency measurement exists yet. The post does not describe how Pro Voice checks that the speaker consented to being cloned.

Analysis: In voice agents, the hard parts are now turn-taking and reading numbers, names and dates correctly, and Mistral's choice of benchmarks says as much. For Indian deployments, the useful news is Indic transcription. On the speech side, Hindi is the only Indic language with a ready-made voice preset. Both launches point the same way: buyers in regulated sectors are starting to choose AI by where it runs and who can access it, and benchmark scores come second. Teams evaluating voice stacks should test them on their own call recordings, with their own number and name formats, before trusting any vendor's chart.