NEMOTRON 3 ULTRA IS HERE
MoE – 550B, 55B Active – Base and Instruct in BF16
– NVFP4 Instruct (350GB)
– 1M Context Window
– Supports MTP Spec Decoding This release IS NOT just about the Weights
It's an Opensource Frontier Intelligence Handbook – Training Data (Pre & Post)
–
INNOVATION
-

NEMOTRON 3 ULTRA: Open Source Frontier Intelligence Handbook with 1M Context
By
–
-
How Executives View AI in 2026: Key Perspectives
By
–
How Executives Are Thinking About AI in 2026
#AI #AIio #AIInnovation #ML #DataScience #Futureofwork @HaroldSinnott @fogoros @iainljbrown @NandoDF @katecrawford @drhassanrashidi @YuHelenYu -
NVIDIA Cosmos 3 launch: download models on HuggingFace and GitHub
By
–
Get started with Cosmos 3: Download Models on @huggingface → https://
huggingface.co/collections/nv
idia/cosmos3
… Customize Models on GitHub → https://
github.com/nvidia/Cosmos -
Cosmos 3: NVIDIA’s open world foundation model for physical AI
By
–
Physical AI needs to understand the world before it can act in it.
— NVIDIA (@nvidia) 4 juin 2026
Introducing Cosmos 3, the open world foundation model and the first omni-model for physical AI.
It understands and generates across text, image, video, sound and action – thanks to a new breakthrough… https://t.co/K5RyHTY7gWPhysical AI needs to understand the world before it can act in it. Introducing Cosmos 3, the open world foundation model and the first omni-model for physical AI. It understands and generates across text, image, video, sound and action – thanks to a new breakthrough
-
AI industry overload: 30k daily posts, Scobleizer uses AI to filter
By
–
Well, it's a tough one because there are about 30,000 posts every day from the AI industry. I built my own AI to read all of that, and it's certainly tough to get seen. This platform is aiming at advertisers of a certain kind, so their algorithm is picking a specific kind of
-
Gary Marcus claims vindication for neurosymbolic AI in LLMs like Code Interpreter
By
–
did you actually read the paper? or just the title? i said you would need neurosymbolic AI to supplement deep learning – and that is exactly what happened with code interpreter, claude code etc. so i was correct!
-
Brex’s $2,000 monthly AI allowance exceeds Uber’s $1,500 per tool
By
–
More generous allowance than Uber then, Brex are allowing $2,000/month, Uber was $1,500 (Though Uber's caps are per-tool and we don't know how many tools they support, if they allow Cursor and Copilot and Claude Code and Codex that's $6,000/month for engineers who use them all)
-
Gary Marcus: Claude code builds tools, creating a neurosymbolic system.
By
–
the system as whole uses symbols (eg in the tools); claude code builds those tools in that’s exactly what it means to be a neurosymbolic system, to use symbols tools in addition to neural networks.
-

Microsoft shares deep dive into data engineering for frontier AI models
By
–
"MAI-Thinking-1: Building a Hill-Climbing Machine" Microsoft just did something almost no frontier AI lab has done before They shared how they engineered the data behind a frontier-scale model in unusual depth. From data collection and eval decontamination, to data mix
-

NVIDIA’s Rafiqspace AI achieves 97.7% Bahasa Indonesia ASR accuracy
By
–
When laws and oversight depend on transcripts, 70–80% isn’t enough. http://
Rafiqspace.ai hit 97.7% Bahasa Indonesia accuracy (2.3% WER) with fine‑tuned Nemotron Parakeet ASR—outperforming global tools while cutting per‑hour costs by up to 90%.