Humanoids
NVIDIA’s GR00T 2 learns robot skills from human video, code due December
NVIDIA says GR00T 2 hits 95% on 50 robot tasks after five demos each. Its own whitepaper says that number is high partly by construction.

Humanoids
NVIDIA says GR00T 2 hits 95% on 50 robot tasks after five demos each. Its own whitepaper says that number is high partly by construction.

Open models
Nous Research's open Arete-27B retunes Qwen3.8-27B. On Nous's own 803-prompt test, propaganda answers fell from 22% to 0.9%. No outside replication yet.

Image and video models
Google made Gemini Omni 1.2 Flash generally available on 5 November 2026, with native 15-second clips and 720p API pricing of $0.07 a second, 30% below Omni 1.1.

Industrial robots
Japan's AIRoA released a bimanual robot model trained only on handheld gripper data, reporting 78% success on ten tasks at an unseen site against 27% for the strongest baseline it tested.

Industrial robots
NVIDIA's VFlash paper reports about 80% lower robot-plan latency on Jetson Thor and DGX Spark versus autoregressive decoding, with author-run ADE gains up to 61%.

AI chips
AMD's data center sales nearly doubled to $8.5 billion as MI455X GPUs shipped. It says AI chips will outsell server CPUs in Q4 but gives no split.

Explainer
Arena, Artificial Analysis, LiveBench and Scale's SEAL boards each test something different. Here is what each one measures and which to check for your use case.

Language models
Meituan's 1.6 trillion parameter model posts coding and agent scores near Claude Opus 5. Every score so far is Meituan's own, and the weights are not out.

Open models
Mistral says its 1-trillion-parameter model leads open-weight coding. Artificial Analysis ranks it second and measured lower scores on two of Mistral's headline tests.
