What changed in AI, and what it means for your work.

Mistral has released Voxtral 3, a family of speech-to-text and text-to-speech models, plus a voice agent SDK built on LiveKit that listens, reasons, calls tools and replies in one streaming pipeline; Mistral says each turn completes in under one second. Transcription covers 50 languages, including the 24 official EU languages and many low-resource Asian and Indic languages, with speaker labels and the option to supply key terms to improve accuracy, and Mistral says its Small model outperforms Whisper, Scribe v2 and Gemini 3.5 Transcribe. Mistral lists transcription from $0.0035 a minute, speech generation at $13 per million characters, and the voice agent API with Mistral Small 4 at $0.04 a minute. It is available through Mistral's API on European, US and global endpoints, or for self-deployment.

Why it matters: If you run a support line or take bookings by phone, a voice agent pilot at about $0.04 a minute is cheap to try; test transcription on your own recorded calls in the languages your customers speak before trusting the benchmark claims.

Mistral has released Timestral, its first family of time-series forecasting models, in four sizes from 5M to 209M parameters, with the weights on Hugging Face. The models forecast from past data without first being trained on yours, give a range of likely outcomes rather than a single number, and can take in factors known in advance, such as holidays or production schedules. Mistral says the largest ranks first on three public forecasting benchmarks, GIFT-Eval, TIME and fev-bench, and that it needs no GPU: a one-day forecast from three weeks of 15-minute readings took 140 ms on an 8-thread server CPU. A technical report is still to come. The Modified MIT licence is free to use, except for companies with more than $20 million in monthly revenue, which need a commercial licence. On Hacker News (362 points, 62 comments), 5 of the 15 top-level commenters welcome the release or more open competition in forecasting. Four are skeptical: one of how the charts are drawn, one of real-world use, one of Mistral in general, and one of whether forecasting can be judged fairly at all. Three raise technical questions, including whether all time series belong in one model and how Timestral compares with Nixtla's TimeGPT. Two are cautiously positive, one defending foundation models when you have thousands of series to forecast and one impressed by the scores but calling the licence restrictive. One comment is about Mistral's coding models, not Timestral. No consensus. Sorted take: Worth a side-by-side test on your own past sales or ops data if you already forecast by hand or with classical models; treat Mistral's leaderboard claims as marketing until you see the results yourself, and check the $20 million monthly revenue licence limit before you ship anything.

Why it matters: If you forecast sales, cash flow, stock levels or call volumes, try the small model on last year's data and compare its forecasts with what actually happened before relying on it.