Models — page 7
Nvidia Blackwell sweeps MLPerf Training 6.0 across all seven benchmarks
GB300 NVL72 delivers 1.6x faster training than GB200 NVL72, and CoreWeave hits a 2.02-minute time-to-train on DeepSeek-V3 671B.
Loft Orbital's Yam-9 runs first vision-language model in orbit
A Google DeepMind Gemma 3 model running on a Nvidia Jetson chip aboard Yam-9 identified targets from natural language queries.
NVIDIA Blackwell runs 20x more agents per megawatt on new AgentPerf benchmark
Artificial Analysis launches the first agentic AI benchmark, and the GB300 NVL72 takes the top spot against Hopper on DeepSeek V4 Pro.
Avataar's Varya undercuts Veo and Runway on price by 20x for Indian video AI
The Peak XV-backed startup distilled Alibaba's Wan 2.2 into a four-step model that generates 720p clips for $0.005 per second.
Anthropic reverses hidden Claude Fable 5 sabotage of AI researchers
After backlash, Anthropic will make its frontier-LLM safeguards visible instead of silently degrading Claude's output for rival AI work.
Anthropic's Claude Fable 5 refuses basic biology questions by design
Anthropic told The Verge Fable's guardrails are 'overly conservative' to block bioweapons queries, routing routine biology asks to Opus 4.8.
Writer researchers find memory tools make AI models more sycophantic, less accurate
Two new papers show that storing user preferences pulls models toward wrong answers, with Mem0 and Zep amplifying the bias.
Google DeepMind's DiffusionGemma generates text 4x faster on GPUs
The 26B MoE model hits 1,000+ tokens per second on an H100 by drafting 256 tokens in parallel instead of one at a time.
Decart launches Oasis 3, a real-time world model for driving simulation
The $4B startup is pricing API access at $0.02 per second and betting developers will build on top of it like the early OpenAI ecosystem.
Anthropic's Claude Fable 5 spins up playable games from a single prompt
Wharton's Ethan Mollick says the new Mythos-class model executed multi-page specs autonomously for up to a dozen hours.
Anthropic releases Claude Fable 5, its first public Mythos-class model
The lab once said Mythos was too dangerous to ship. New safeguards route 5% of risky prompts back to Opus 4.8.
Google DeepMind launches Gemma 4 12B, an encoder-free multimodal model for laptops
The mid-sized open model runs on 16GB of VRAM and nears 26B benchmark performance, with native audio and vision baked into the LLM backbone.
Google DeepMind ships Gemini 3.5 Live Translate across 70+ languages
The new audio model streams speech-to-speech translation continuously, lagging the speaker by just a few seconds across Meet, Translate, and the Live API.
Anthropic logs 8x code merge jump as researchers benchmark AI gaming society's rules
Anthropic sees early signs of recursive self-improvement, a new benchmark tests AI loophole-hunting, and RL drones beat a human champion at 22 m/s.
ECMWF's AIFS runs forecasts on 1,000x less energy than physics-based model
Machine learning is rewriting weather forecasting at a fraction of the compute cost, but climate modeling won't surrender to neural nets.
Estonia's new benchmark ranks Claude Opus 4.7 best at resisting Russian propaganda
The Estonian Language Institute scored dozens of LLMs across 14 propaganda categories; Anthropic took six of the top 10 spots.
Google releases Gemma 4 12B, sized to run locally on a 16GB laptop
The new mid-weight Gemma slots between mobile and workstation variants, with model weights just under 18GB available on Hugging Face and Kaggle.
Nvidia and Unitree pair Thor chip with H2 Plus humanoid in US-China robot blueprint
Jensen Huang's blueprint puts a Thor T5000 brain inside a 6-foot, 150-pound Chinese-built body — and undercuts rival humanoids on price.
Nvidia unveils GraspGen-X, LCDrive and NitroGen foundation models at CVPR
Three new physical AI models target zero-shot robot grasping, faster autonomous-vehicle reasoning, and gameplay-trained embodied agents.
Nvidia ships physical AI agent skills at CVPR, anchored by 32B Alpamayo 2 Super
New agent skills built on Cosmos 3 automate scene reconstruction, simulation and policy training for AVs, robots and vision AI.
Microsoft to unveil MAI-Thinking-1 reasoning model and Copilot super app at Build
Build 2026 lands in San Francisco June 2 with new in-house models, a Windows 11 developer mode, and an RTX Spark push.
Windborne's WeatherMesh 6 out-forecasts ECMWF with 400 balloons feeding the model
The Stanford-founded startup says its sixth model is as accurate five days out as a traditional forecast is the day before.
Intel's Crescent Island AI chip ships this year, undercutting Nvidia on cost and cooling
The air-cooled inference GPU uses LPDDR5 memory instead of HBM, betting that cheaper silicon wins the inference market Nvidia hasn't locked down.
OpenAI model disproves Erdős unit distance conjecture after 80 years
An internal OpenAI system produced the first proof resolving a major open conjecture in discrete geometry, drawing praise from Fields Medalist Tim Gowers.
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