Frontier AIArchive
Stories older than 30 days leave the latest homepage and land here. Original pieces, summaries, and links out to external reporting stay intact — nothing is deleted. Use Search to find any story across latest and archive.
On 12 August 2026 xAI shipped Grok 4.6 for long-running agents, coding, and knowledge work. Lab table ties GPT-5.6 Sol Max on Artificial Analysis Intelligence Index at $2/$6 per million base tokens, live in Cursor, Grok Build, Grok Bot, and the API.
On 11 August 2026 NVIDIA released Nemotron 3.5 Lightning, an open 30B mixture-of-experts model with 3B active parameters for high-volume always-on agent execution under OpenMDW-1.1, plus NeMo Switchyard for multi-model routing.
Meta Superintelligence Labs released Muse Glimmer on 10 August 2026 as Apache 2.0 open weights: about 29.6B parameters with a perception encoder, distilled from Muse Spark for always-on local agents on one consumer GPU. Hub id is meta-models/Muse-Glimmer-30B. Size-class peer names and agentic scores stay Meta-reported.
Preliminary evals of OpenAI's upcoming Astra cannot rule out Critical cybersecurity under the Preparedness Framework. OpenAI is treating it as its first Critical cyber model, tightening controls, and still aiming for broad defender access.
On 6 August 2026 Google DeepMind published a Nature paper on WeatherNext cyclone forecasting and open-sourced WeatherNext 2 and WeatherNext Cyclones code plus weights. The lab reports about a day of extra lead time on track, intensity, and wind structure in its own evaluations.
On 4 August 2026 NVIDIA put Alpamayo 2 Super on Hugging Face for commercial use: a 34B vision-language-action model for robotaxi and AV development under OpenMDW-1.1. Blog and Hub card pair a 32B Cosmos 3 Super reasoner with a 2.3B diffusion action expert and vendor LingoQA and driving metrics.
Mistral released Shieldstral on 4 August 2026 as Apache 2.0 open weights: a 3B multimodal safety classifier that takes plain-language policies at inference and returns a continuous yes/no safety score. Hub card is mistralai/Shieldstral-1.0-3B. Vendor tables are competitive with much larger open guards on several text axes and lead VLGuard and UnsafeBench at the stated threshold.
Qwen's 2 August blog puts Qwen3.8-Max on QwenCloud at 2.4T parameters (95B active) and says Max-class open weights land next week. Hosted is live; weights are still a calendar promise.
On 31 July 2026, Alibaba's Qwen account launched Qwen-Audio-3.0-ASR-Flash with context-aware ASR, domain terms, hotwords, and structured polishing. Model Studio lists API ids qwen-audio-3.0-asr-flash and filetrans siblings. Medical and industrial recall figures are vendor internal tests.
On 30 July 2026 Google DeepMind launched Gemini Robotics 2 as a three-model family: a full-body vision-language-action stack, embodied reasoning ER 2, and an on-device VLA. ER 2 is in API and AI Studio preview; the VLA and on-device models stay early-access partner only.
Showing 10 of 20 stories
Free
Free daily email of model releases, benchmarks, and lab notes from the last 24 hours across OpenAI, Anthropic, DeepMind, Meta, NVIDIA, and more. Capability-first, not company drama.
Opens Buttondown to finish signup. Confirm by email if asked. No ads. Unsubscribe anytime.
Looking for something recent? Live desk · Morning brief · Search.
