NVIDIA의 오픈 모델 서한과 실시간 음성 코딩·124B MoE·로컬 대형 모델 실험

Jensen Huang

@JensenHuang

For my first post, I’m sharing a letter @NVIDIA signed on why open models matter. AI will transform every industry, power every company, and be built by every country. Open models strengthen safety and cybersecurity, accelerate innovation and diffusion, and enable sovereignty. The world needs both frontier closed models and frontier open models. https:// images.nvidia.com/pdf/Open-Weigh ts-and-American-AI-Leadership.pdf …

Codex just got real time voice mode. Talk to your coding agents like coworkers. Talk to them. Interrupt them. Send them to work.

Free 124B MoE that runs like a much smaller model but thinks like a giant 124B parameters, Only 5.1B active per token. production agent Ant Group’s new Ling-3.0-flash matches or beats their own 1T flagship on most benchmarks with 1/8 the total params and 1/12 the active ones. Hybrid KDA, MLA attention, native 256K context (scales to 1M), and some of the cleanest agent demos I’ve seen: - Built a full 3D city in Blender from one prompt - Ran an entire autonomous research team (hypothesis to paper then slides) - Handled real Office workflows end-to-end - Generated working web apps and design systems from code alone Free on OpenRouter until August 3.

Crazy, Local AI crossed another threshold: frontier GLM-5.2 running on 3× DGX Sparks at home with 248k context. GLM-5.2 running on 3× NVIDIA DGX Sparks at home: - 248k token context - 15-20 tok/s (content-dependent) - NVFP4 + AQLM hybrid quantization - MTP speculative decoding - Locally 744B MoE model. - Custom vLLM fork + TP3 Memory is extremely tight, yet they’re delivering usable speed. One of the cleanest local inference setups I’ve seen: This is exactly the kind of practical frontier work that moves local inference forward.

the setup and repo are already public. https:// github.com/MiaAI-Lab/GLM- 5.2-NVFP4-AQLM-Triple-DGX-Sparks … The full orchestration scripts, Docker setup, and Ray cluster config are open-sourced. For anyone running large models locally or building long-horizon agents this best for you

Pocket Analyst Tool (PAT) is an example of how we're combining decades of codified investment knowledge with large language models, agentic workflows, and modern software architecture to build AI tools for investment research. Watch how we built PAT.

NVIDIA will prepay Amkor $1.5 billion under a multi-year deal to expand advanced chip packaging and testing in Arizona. The agreement targets the packaging bottleneck for AI accelerators and includes joint development of packaging for NVIDIA's next-generation AI platforms.