DeepMind open-sources WeatherNext 2: 15-day forecasts on one TPU
Open weights and code on GitHub, cyclone lead time up a full day. Also: Kimi K3 in Copilot, code-review effort dials, Antigravity 2.6, faster long-context.
Copy markdown
Grab the weights, build without a weather API
Google DeepMind open-sourced WeatherNext 2's code and model weights on GitHub — a 28-km flagship and a 111-km "mini" that runs a full 15-day forecast in under a minute on a single TPU. You can now build weather-aware apps (logistics, insurance, agriculture, event planning) on your own infra instead of paying per call to a hosted weather API.
A full day more cyclone warning, 1,000-member ensembles
One network predicts storm track and intensity and scales to 1,000-member ensembles (up from ~50 in prior runs), giving forecasters roughly a day more lead time than leading operational models. It helped anticipate Hurricane Melissa's rapid intensification during the 2025 Atlantic season; the underlying research published in Nature on Aug 6.
Kimi K3 is now selectable in GitHub Copilot
Moonshot's 2.8-trillion-parameter Kimi K3 — an open-weight model that leads Arena.ai's frontend-code arena — is now available in GitHub Copilot, billed at provider list pricing (Moonshot lists roughly $3/$15 per million input/output tokens). If you're already on Copilot you can switch to it today without wiring up a separate provider key.
Copilot code review gets Lite and Balanced dials
GitHub made Lite and Balanced effort levels generally available for Copilot code review, letting you trade review depth for speed and cost per change. Run Lite on low-risk diffs and reserve the heavier pass for complex PRs.
Antigravity 2.6 speeds up long agent sessions
Google's Antigravity IDE shipped 2.6.0 with faster long conversations and more reliable hooks and subagents, plus 21 bug fixes across chat, comments, and the file viewer. Worth updating if your agent runs bog down on long threads.
GPT-5.6 Fast mode now takes 272K-token prompts
OpenAI's API changelog says Fast mode now supports long-context requests for GPT-5.6 Sol, Terra, and Luna — prompts over 272K tokens run up to 2.5x faster than the Standard tier. Whole-repo reviews and long transcripts no longer force you off Fast mode.