Forecast from ordinary servers
Mistral released Timestral, a family of models for forecasting measurements that change over time. It handles related data series together, produces probability ranges rather than one prediction, and can incorporate known future factors such as weather forecasts or production schedules.
The four models range from 5 million to 209 million parameters and run without a GPU. Mistral says the largest can forecast one day from three weeks of 15-minute readings in 140 milliseconds on an eight-thread server CPU. A public GitHub repository lets builders run Timestral on their own data, while Mistral Forge supports fine-tuning and deployment on customer hardware, including air-gapped systems.
Speech and decisions in one model
IBM’s Granite Switch 4.1 combines conversation, speech transcription, and specialized decision models behind one hosted endpoint. Its 8-billion and 30-billion parameter versions can move from transcribing audio to summarizing it or checking user-defined flags.
IBM says the combined components use about as much GPU memory as the base model alone. Builders can compose their own Switch models from task-specific adapters, although an adapter training kit is still planned for later.
One cart across Google
Google’s Universal Cart is rolling out in the U.S. across Search and the Gemini app. It collects products from multiple retailers, watches for price drops and restocks, surfaces promo codes, and can suggest a payment method using card benefits in Google Wallet.
Checkout either uses Google Pay at participating stores or transfers items to the merchant’s site. The merchant remains responsible for the sale, confirmation, and customer support.
Also
- Score a video automatically ElevenCreative generates music and timestamped sound effects from uploaded footage, then places them on separate editable layers.
- Find sources sooner in AI Search Google is placing inline links higher in AI responses and testing a dedicated section for relevant official pages.
- Give outside evaluators deeper access Anthropic has embedded six employees from METR, a nonprofit that evaluates frontier AI models, with internal access and independent publication rights subject to narrow predefined redactions.
- Assign security alerts to Copilot GitHub’s agentic autofix explores code, proposes a repair, reruns CodeQL, and opens a draft pull request. It requires eligible Enterprise Cloud security products, admin enablement, Copilot billing, AI Credits, and Actions minutes.
Pass it on