A quieter day on the surface, but three releases show where the frontier is actually moving: weather and speech get cheaper and sharper, and open-weight labs keep shipping bigger models.
Google DeepMind ships WeatherNext 3 â satellite-first global forecasting
Google DeepMind and Google Research launched WeatherNext 3, its sharpest AI weather model yet, now live in Google Search, the Gemini app, Maps, the Maps Platform Weather API, and Earth Engine.
- Trained directly on live geostationary satellite mosaics instead of Numerical Weather Prediction (NWP) output, cutting data lag from ~7 hours to 3â4.
- New global forecast every hour at up to 5 km resolution â about 5Ã sharper than WeatherNext 2's 25 km / 6-hour grid.
- Up to 60% better precipitation CRPS versus IMERG, plus new turbine-height wind and solar-radiation variables aimed squarely at renewable-grid operators.
- Runs a 64-member ensemble for probabilistic output and adds station-level predictions conditioned on local terrain.
The angle: AI is leaving the chat box and quietly becoming infrastructure for energy dispatch.
Tencent open-sources Hunyuan Hy4 Preview â 770B MoE, 1M context
Tencent released and open-weighted Hy4 Preview under Apache 2.0 (Hugging Face, ModelScope, GitCode, CNB), its largest open model to date.
- 770B total / 49B active parameters (Mixture-of-Experts, 78 layers, 256 routed experts + 1 shared per layer), 1M-token context.
- Trained with what Tencent calls a recursive self-improvement loop â the model optimized parts of its own training, data, and low-level operators, lifting end-to-end throughput ~31.8%.
- Benchmarked 8th on Code Arena WebDev (up from Hy3's 34th) and 64.3 on DeepSWE, ahead of Qwen3.8-Max (56.6) and DeepSeek-V4 Pro (62.7).
- In a 163-expert blind eval on 203 engineering tasks it edged GLM-5.3 (2.99 vs 2.92) and Kimi K3 (2.99 vs 2.94) â a margin inside the noise, so judge it on price and license, not a 0.07-point lead.
- API at $0.834 in / $2.501 out / $0.042 cached per M tokens; free on WorkBuddy and CodeBuddy for two weeks. Text-only for now.
The angle: the open-weight frontier has effectively converged â differentiation is now price, license, and serving footprint.
Microsoft MAI-Transcribe-2 â 5.2% WER, 10Ã faster, $0.10/hour
Microsoft AI launched MAI-Transcribe-2, which it claims is the fastest, most accurate, and cheapest speech model available.
- 5.2% average WER on the FLEURS benchmark across 60 languages (2.0% in non-streaming mode), ranking first on FLEURS and second on the Artificial Analysis WER leaderboard.
- Up to 10Ã faster than OpenAI's GPT-Transcribe, 7Ã faster than ElevenLabs Scribe v2, 5Ã faster than Gemini 3.5 Transcribe â roughly an hour of audio back in ~10 seconds.
- Ships with speaker diarization, word-level timestamps, keyword biasing, automatic language ID, code-switching (Hinglish/Spanglish), and verbatim/clean styles â diarization included, not a paid add-on.
- Promotional price $0.10 per audio hour through end of 2026; available via Microsoft Foundry, MAI Playground, and OpenRouter.
The angle: transcription just joined the "good enough, cheap, and fast" tier that text generation is racing toward.
More daily AI briefings at AI Nexus Daily.
Top comments (0)