NVIDIA Releases Multimodal MoE Model "Nemotron-3-Nano-Omni-30B"
NVIDIA has released a new multimodal large language model, "Nemotron-3-Nano-Omni-30B," on Hugging Face. It adopts a hybrid MoE (Mixture of Experts) architecture that fuses Mamba2 and Transformer, with a sparse design that activates roughly 3B parameters per token out of 31B total.
Input covers four modalities—video, audio, image, and text—with text as the output. The maximum context length is 256K tokens, and the model targets a broad range of enterprise use cases including video analysis, speech transcription, OCR, document intelligence, and GUI operation. Reasoning mode is enabled by default, and both thinking and instruct modes can be toggled.
Notably, the NVFP4 quantized version fits within 21GB and runs on the RTX 5090 (32GB) and Jetson Thor. Commercial use is permitted under the NVIDIA Open Model Agreement. That said, the fact that external models such as Qwen3-VL and Qwen3.5 were used in training hints at the ecosystem dependencies behind open models.
White House Considers Pre-Release Review System for AI Models
According to The New York Times, the White House is considering introducing a system under which the government reviews AI models before they are publicly released. The mechanism would evaluate safety and risks before a model ships, and it could significantly reshape how regulation works.
The topic is also being actively debated in the LocalLLaMA community, where concerns are being raised about the impact on open-source models. The scope and criteria for review remain unclear at this stage, and many comments point to potential harm to free expression and innovation. No specific legislation or implementation timeline has been revealed yet.
EU's €20B AI Computing Investment Criticized as Waste
According to Politico, criticism is mounting that the EU's roughly €20 billion investment in AI computing infrastructure amounts to "money poured into a pipe dream." Despite spending enormous sums on supercomputers and cloud infrastructure, actual utilization has been limited and the competitiveness gap with the US and China has not narrowed, critics say.
The EU's AI strategy builds on cloud sovereignty initiatives such as "Gaia-X," but a fundamental problem looms in the background: the bloc has failed to keep pace with the private sector's rapid investment cycles. Structural issues have also been pointed out—data centers are being built, but the AI ecosystems and talent development needed to make use of them are not keeping up.
Image Generation AI Drives App Growth, Outpacing Chatbot Updates
TechCrunch reported on an analysis by Appfigures showing that image generation AI model launches generate 6.5 times more download growth than chatbot feature updates.
However, few apps have managed to turn this download growth into revenue. Visual novelty attracts users' attention, but in many cases it does not lead to sustained usage or paid conversion. The findings once again highlight the persistent gulf between "attention-grabbing power" and "monetization" in the AI app market.
The Limits of AI Agents — Engineers Are Moving Back to Deterministic Systems
A Hacker News post asking engineers who have experienced "the move from AI agent loops to simpler deterministic systems" drew significant attention. The poster pointed to the challenges that autonomous agent loops bring—latency, non-determinism, accuracy, and cost—and asked at what stage of the product lifecycle they decided that "the agentic approach isn't appropriate."
AI agents are attractive as a concept, but an increasing number of teams are running into predictability and maintainability problems in real-world operations. Especially in areas where production-grade reliability is required, many voices are gathering around the view that traditional pipeline approaches remain the pragmatic choice. It is an interesting indicator of the gap between the trend and practical reality.
JuliaHub Raises $65M to Compete with MathWorks' Simulink
According to Axios, AI startup JuliaHub has raised $65 million. Led by Bob Muglia, former head of Microsoft's product division, the company is pushing the commercialization of Julia, a programming language for scientific and technical computing, and is positioned as a direct competitor to MathWorks' Simulink.
The Julia language has spread through the research community on the promise of being faster than Python and easier to write than C, but commercial adoption has remained limited. This funding round can be seen as a strategic move targeting the intersection of AI, simulation, and hardware engineering.