Models — page 3
Ex-Meta scientists launch Perceptron with $21M for factory-floor vision AI
Perceptron's open-weight Isaac 0.5 model targets warehouses and factories, trained on a million hours of video plus petabyte-scale robotic trajectories.
IBM ships Granite 4.2 with 3B, 8B, and 30B variants for local deployment
The reasoning-focused release adds agentic RL training on the larger variants and a 128,000-token context window across the family.
Anthropic merges Claude and Claude Cowork memory into one system
Claude will now carry context across chat and Cowork, and users can read, edit, or delete stored memories on any topic.
Apple's M6 Mac mini and M5 Ultra Mac Studio target local AI inference
Apple's 2nm M6 and dual-die M5 Ultra push up to 512GB of unified memory and 1.2TB/s bandwidth — aimed squarely at developers running local LLMs.
OpenAI's Jalapeño inference chip beats Nvidia Blackwell on SemiAnalysis benchmark
The Broadcom-partnered chip ships in small volumes at the end of 2026, with broader deployment in 2027.
Nvidia ships Groq 3 LPX to accelerate Vera Rubin agentic inference
The new accelerator hits 3,400 output tokens per second on 100,000-token contexts, 4x the nearest platform, with Nebius as first cloud adopter.
Ox Alpha stealth model lands on OpenRouter with no owner attached
A free reasoning model appeared on OpenRouter this week with no named creator, drawing praise from Stripe's Patrick Collison and speculation across the AI community.
Inherent's Faraday agent beats Claude and GPT-5.5 at replicating research on a 27B model
The London lab, fresh off a $50M seed, says its DeepMind-alumni-built agent matches frontier systems using a fraction of the parameters.
Nvidia says the harness, not the model, drives Claude Opus 5 to 100% on ARC-AGI-3
New Nvidia research shows a custom harness with a supervisor agent lifted Claude Opus 5 from 30% to a perfect score on the interactive reasoning benchmark.
DeepMind's SIMA 2 agent plays, reasons and learns in 3D game worlds
Google DeepMind extends 15 years of games research from Atari to EVE Online with a new agent that collaborates with human players.
Generalist AI's robots improvise chores with 59% task success
The Cambridge startup founded by ex-DeepMind and Boston Dynamics researchers is chasing a GPT-3 moment for physical intelligence.
DiG-bench and Faraday mark AI's slow climb toward recursive self-improvement
New benchmarks show frontier models cracking discovery games and replicating research papers, with Opus 5 and a 27B scientist model leading the way.
Google makes visible watermark on AI images, video, and audio optional
Nano Banana, Omni, and Lyria outputs can now ship without the visible tag; invisible SynthID and C2PA metadata stay on.
OpenAI launches Ultrafast mode, pushing GPT-5.6 Sol to 750 tokens per second
The new preview mode runs at 14x standard speed, powered by a Cerebras partnership, and targets enterprise workflows where latency is the constraint.
Google ships Gemini 3.7 Flash at half the price of 3.6, three weeks later
The new workhorse model posts 43.6% on FrontierCode 1.1 and 65.3% on DeepSWE v1.1, with input tokens at $0.75 per million.
Google DeepMind ships sign language AI to Pixel 11 with new SL2T model
SL2T translates ASL to English on-device in Gboard and Live Transcribe, trained on 100,000+ hours across 50+ sign languages.
Anthropic's unreleased model advances the Riemann hypothesis after 31M tokens
Coordinating 60 subagents across a day and a half, the model tested 650 ideas and improved the known bound on a 150-year-old problem.
OpenAI's Astra model solves 10 long-open math problems for $2,000 in tokens
The unreleased model cracked problems that had eluded mathematicians for decades — and set off a credit dispute with the researchers whose work it built on.
Nvidia releases Nemotron 3.5 Lightning and NeMo Switchyard for agent workloads
The 30B mixture-of-experts model runs 4x faster on output, while the routing library cuts task cost to a third of Opus 4.8.
Nvidia ships Nemotron 3.5 Lightning, a 30B open MoE for local agents
The new open-weights model runs 4x faster than class rivals and slots into RTX PCs, DGX Spark, and Jetson for always-on agentic workloads.
Meta pivots back to open weights with Muse Glimmer and a 6,000-word Zuckerberg essay
A 30B open-weight model, a promise to open Muse Spark 1.2, and a manifesto against 'singular superintelligence' mark Meta's latest AI reset.
Meta releases Muse Glimmer, a 30B open-weight model for on-device AI agents
The Apache 2.0 model runs locally on a single consumer GPU and sketches Zuckerberg's 'personal superintelligence' pitch.
Intology's Locus beats human baseline on PostTrainBench, signaling AI R&D automation
Locus hits 51.6% on PostTrainBench+ with 4,000+ H100 hours, surpassing the 51.1% human baseline and every frontier-agent competitor.
DeepMind's WeatherNext gives hurricane forecasters an extra day of lead time
The open-sourced AI model predicted Hurricane Melissa's Category 5 landfall in Jamaica five days out with 80% confidence.
Stay ahead of everyone in AI.
The tightly edited AI news email engineers, founders, and investors actually open. One email. Every weekday. Five minutes to finish.
The briefing read inside teams at