Google DeepMind and Research release WeatherNext 3, global AI weather model with hourly forecasts and 5-km resolution
On September 3, 2026, Google DeepMind and Google Research released WeatherNext 3, described as the most advanced global AI weather model to date. The model ingests real-time satellite observations, regenerates global forecasts hourly, and improves spatial resolution for some surface variables to about 5 kilometers. This builds on prior AI weather models like GraphCast, which generated 10-day global forecasts in about 1 minute and outperformed ECMWF's HRES on over 90% of evaluation metrics at the time.
Coverage timeline
机器之心ScienceAI
编辑丨% 每天早晨,人们出门之前通常只关心一个问题:今天到底会不会下雨。 但对于农业、电力调度、航空和防灾来说,天气预报远不只是「晴还是雨」。一场雷雨提前几十分钟出现,一片云层移动几公里,或者风速突然发生变化,都可能改变接下来数小时的决策。 不过大气本身就是一个变化极快、尺度极大的复杂系统。传统天气预报依靠数值天气预报(NWP),通过超级计算机求解描述大气运动的物理方程,再不断更新未来天气状态。这套方法已经相当成熟,却也意味着巨大的计算成本。 过去几年,谷歌 DeepMind 开始尝试用 AI 走另一条路。 从 GraphCast 到 GenCast,再到 WeatherNext,Google 一直在探索一个问题: 如果让 AI 直接从大量天气数据中学习大气如何变化,天气预报能不能变得更快、更细,而且覆盖整个地球? 现在,答案又向前推进了一步。2026 年 9 月 3 日,谷歌 DeepMind 和谷歌 Research 发布了 WeatherNext 3,并称其为目前最先进的全球 AI 天气模型。它最明显的变化并不是简单地「换了一个更大的模型」,模型能够直接吸收 实时卫星观测 ,每小时重新生成全球预报,并将部分地表变量的空间分辨率提高到约 5 公里。 原文链接: https://blog.google/innovation-and-ai/models-and-research/google-deepmind/introducing-weathernext-3/ 以前所未有的分辨率快速预报 传统 NWP 系统的优势在于物理机制清晰,但它需要巨大的计算资源。与此同时,全球天气数据也并不是随时、随地、以相同精度存在的。而这给了谷歌一个切入口。 2023年,谷歌 DeepMind 发布 GraphCast 时,AI 天气预报第一次大规模展示出令人印象深刻的速度优势。GraphCast 能够在约 1 分钟内生成未来 10 天的全球天气预测,而传统高水平数值预报系统则需要超级计算机运行数小时。它在当时的测试中,在超过 90% 的评估指标上优于 ECMWF 的高分辨率预报系统 HRES。 而这次推出的 WeatherNext 3 能以多种空间分辨率生成每小时预报,保持从全球宽广风向模式到局部地形的物理一致性。 具体来说,它能以 5 公里的分辨率可视化关键地表变量——如温度和湿度

Related stories
Google AI Model WeatherNext Cyclones Achieves SOTA Tropical Cyclone Forecasts
Google and multiple meteorological agencies published WeatherNext Cyclones (WN-C), an AI model for tropical cyclone forecasting, featured on the cover of Nature. WN-C generates state-of-the-art ensemble forecasts for global tropical cyclone tracks, intensity, and size over 15 days, trained on global analysis data and historical cyclone records. It bridges the gap between global model track forecasts and high-resolution regional model intensity forecasts, with track forecasts exceeding the best global models by over one day and intensity forecasts outperforming specialized models like HAFS, despite using coarser input resolution.
1 reports · 1 sources

Google DeepMind's WeatherNext AI model improves cyclone forecasts by one day
Google DeepMind published a paper in Nature on August 6, 2026, showing that its WeatherNext AI model achieves state-of-the-art accuracy in predicting tropical cyclone track, intensity, and wind structure. The model, trained on nearly 20TB of global atmospheric data and historical cyclone records, provides forecasters with an extra day of predictive accuracy on average, and is now open-sourced.
2 reports · 2 sources

Google launches Gemini 3.8 Flash and Cyber variant, three weeks after 3.7 Flash
Google released Gemini 3.8 Flash, the third Flash update in three months, alongside a cybersecurity-focused Gemini 3.8 Flash Cyber variant. The model improves reasoning and code abilities, with introductory pricing of $0.75 per million input tokens and $3.75 per million output tokens until December 31. Google warns the model may use more tokens at higher effort levels, potentially increasing costs.
8 reports · 6 sources1154 points · 662 comments

Z.ai debuts GLM-5.3 with post-training scaling gains, open weights in two weeks
Z.ai (Zhipu AI) announced GLM-5.3, an open-source model built on the same base as GLM-5.2 with substantially extended post-training, yielding a 50% gain on internal coding benchmarks and state-of-the-art results among open models on Terminal Bench 3.0 and Agents' Last Exam (CLI). The model matches Mythos 5 in white-box code review and vulnerability discovery, and scored 60 on the Artificial Analysis Intelligence Index, tying Kimi K3. Model weights will be released in two weeks after security hardening; API pricing is set at $1.40 per million input tokens and $4.40 per million output tokens.
9 reports · 6 sources1154 points · 573 comments

News updates every hour. Subscribe in your RSS reader to get new stories automatically.
Subscribe