WeatherNext 3 (DeepMind / Google Research, Sep 3 2026)

Generated by Grok Bot research on 2026-09-03. WebSearch + fetch ladder + native X. Treat as raw material — review before promoting.

Filer (Alfred): /clipping-file → /clipping-promote into vault/threads/artificial-intelligence → /ingest-pending for this source only. Root + one follow-up X posts are video (x_video: true); text claims below are hydrated from the blog and developers docs — promote those. Do not treat video demos as grain until /transcribe-clipping. Do not re-file same-day Gemini 3.8 Flash, K2 Horizon, Qwen, Mostik, or Fable 5.1 permalinks. Skip outage chatter and the OpenAI video-only post (https://x.com/OpenAI/status/2095527557924082061) until transcribed. Atlas (https://x.com/theworldlabs/status/2094839756329041984, Sep 1) is older than today's X pass window; leave it out of this file.

Summary

On Sep 3, 2026 Google announced WeatherNext 3 as DeepMind + Google Research's flagship global weather AI: live geostationary satellite mosaics as a direct input, hourly refresh, multi-resolution output down to ~5 km for station-trained temperature/dew point, and product rollout into Search, Gemini, Maps, Maps Platform, Earth Engine, plus developer access via BigQuery / Earth Engine / GCS. Official X from @GoogleDeepMind matches those claims. Numbers disagree slightly across surfaces (blog CRPS vs product "50% more accurate" vs X "50% reduction in error" vs docs Brier/CRPS vs IMERG) — keep them separated, do not flatten.

Findings

Product announcement (Sep 3)

The Keyword / DeepMind blog post dated Sep 03, 2026 states WeatherNext 3 learns from real-time observations, uses raw satellite data to produce a forecast every hour at high resolution, and claims it is "the most advanced and accurate global weather model to date, according to independent live evaluations by Brightband." Product surfaces named: Google Search, Gemini app, Google Maps, Google Maps Platform Weather API, Google Earth Engine, starting that day. Developer path named: BigQuery, Earth Engine, bulk download from Google Cloud Storage.

Architecture and resolution (issuer)

The blog describes a Functional Generative Network (FGN) mesh transformer that ingests live 1-hour geostationary satellite mosaics plus traditional historical analysis, and outputs dense gridded fields, discrete cyclone tracks, and station-level sparse coordinates. Resolution claimed vs WeatherNext 2: surface temperature/moisture visualized at 5 km, other surface vars at 10 km, atmospheric vars (e.g. wind) at 25 km, described as "roughly five times sharper" than WeatherNext 2's 25 km / 6-hour grid.

The developers.google.com WeatherNext 3 model card (fetched) gives operational specs:

Paper PDF linked from the model card: https://storage.googleapis.com/deepmind-media/papers/weathernext_3.pdf (URL present on docs page; PDF body not fetched this pass — cite the link, do not invent paper abstract claims).

Precipitation and clean-energy variables

Blog medium-range precip evaluations (CRPS): up to 60% vs IMERG, 30% vs MRMS, 10% vs rain gauges for early lead times. Same blog's consumer product claim for day-or-more-ahead planning: up to 50% more accurate precipitation forecasts, largest gains where forecasts were historically less reliable.

Docs precip line (models page): trains against ECMWF reanalysis, NASA IMERG, and Google's satellite-radar precip reanalysis; "up to a 50% reduction in Brier score and CRPS compared to numerical weather prediction baselines when evaluated against IMERG observations."

Clean energy (blog): 100-meter wind (turbine-height), high-res cloud cover, sun radiation / solar irradiance for solar farms. Docs list wind_speed_100m, SSRD/FDIR solar components, and cloud-layer fractions on the 0.1° grid.

Official X thread (primary discourse; video flagged)

@GoogleDeepMind root, 2026-09-03 15:03 UTC (video): WeatherNext 3 with @GoogleResearch; learns from real-world real-time observations.

Follow-ups same minute:

• Hourly refresh vs six-hour traditional
• Station training + 5× temp resolution 25 km → 5 km (video)
• Up to 50% reduction in precip error
• Powers Search / Gemini / Maps / Maps Platform; BigQuery, Earth Engine, GCS — links https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-weathernext-3/

Author-adjacent pointer (not lab primary): @FerranAlet claims best global model on Brightband's live leaderboard and links https://owb.brightband.com/. Treat Brightband ranking as issuer/advocate claim until a dated leaderboard snapshot is filed; do not invent ranks or scores.

Same-day X noise intentionally not filed here

• Multi-vendor outage chatter (ChatGPT / Claude / Grok) — ops, not model grain.
• @OpenAI video-only — no text claim; x_video; hold for transcription.
• Karpathy RT of World Labs Atlas (Sep 1 video) — outside "newer than today's X clip" primary window for this 1pm pass.
• Already-filed today: Gemini 3.8 Flash / Cyber, K2 Horizon, Qwen / Mostik open-model pass.

Contradictions and open questions

• Date label: docs model card says Release Date August 2026; public blog + DeepMind X are 2026-09-03. Prefer Sep 3 for "announced / productized"; keep August as the docs table value.
• "50%" is not one number: blog product copy ("50% more accurate precip" for users), X ("50% reduction in error"), docs ("50% reduction in Brier score and CRPS vs NWP baselines on IMERG"), and blog research CRPS (60% / 30% / 10% vs IMERG / MRMS / gauges). Do not collapse into a single 50%.
• Brightband "best / most accurate": blog + Ferran Alet assert it; live board at owb.brightband.com was not snapshot-captured this pass. Flag as contested until a dated board scrape or Brightband primary is filed.
• Paper PDF URL is known; contents not fetched. Do not invent paper numbers.
• Videos on the DeepMind root and 5 km-resolution posts are untranscribed — demo only until /transcribe-clipping.

Provenance

Method: Grok Bot / WebSearch / fetch ladder / native X (search_news + get_users_posts + get_posts_by_ids) Generated: 2026-09-03 Rounds: 1 of 3 — early-exit: single new lab launch with hydrated primary; other same-day X signal was duplicate, outage, or video-only URLs fetched: 2 successful (blog via curl; developers model card via WebFetch), 1 timed out (deepmind.google discover blog index via WebFetch — unused) Fetch notes: blog.google/.../introducing-weathernext-3/ WebFetch not attempted after discover timeout; curl with browser UA succeeded. developers.google.com/weathernext/guides/models WebFetch succeeded. PDF not fetched. Brightband board not fetched. Grok news-cluster summaries used only as discovery pointers — not cited as posts.

Web sources:

• Introducing WeatherNext 3 (Google Keyword / DeepMind, Sep 03 2026) — product announce: hourly satellite, 5 km, Brightband claim, CRPS 60/30/10, 50% consumer precip, Search/Gemini/Maps/Cloud
• WeatherNext 3 model card (Google for Developers) — operational specs, FGN, ensemble 64, precip Brier/CRPS 50% vs NWP on IMERG, paper PDF URL
• WeatherNext 3 PDF (linked from docs) — URL only; body not fetched

X sources:

• X post by @GoogleDeepMind (2026-09-03) — root announce + video
• X post by @GoogleDeepMind (2026-09-03) — hourly vs six-hour
• X post by @GoogleDeepMind (2026-09-03) — station training + 25→5 km (video)
• X post by @GoogleDeepMind (2026-09-03) — up to 50% precip error reduction
• X post by @GoogleDeepMind (2026-09-03) — product surfaces + blog link
• X post by @FerranAlet (2026-09-03) — Brightband live-leaderboard claim (advocate)

Grokipedia:

• not used