要点速览
- NVIDIA releases Nemotron 3 Ultra, a 550B MoE open model with 1M token context, 5x faster inference and 30% lower costs for agent tasks
- Google launches Gemma 4 12B, a encoder-free multimodal model with audio/image input support and 256K context window
- Hugging Face launches LeLab and updates TRL for agent training with trace data
- OpenAI rolls out ChatGPT memory system and updates GPT-Rosalind for life sciences
一、Major Model Releases & Updates
- NVIDIA officially launched Nemotron 3 Ultra, a 550B total parameter, 55B activated parameter MoE open model designed for long-running agents. It delivers 5x faster inference and up to 30% lower costs for complex agent tasks, supports 1M token context, MTP speculative decoding, and is available on Hugging Face Hub with OpenMDW-1.1 license. It has achieved SOTA performance on MMLU, code, and long-context benchmarks, and is supported by transformers, vLLM, SGLang and other frameworks. — via 1 2 3
- Google released Gemma 4 12B, an encoder-free unified multimodal model that supports direct audio and image input without additional encoders. It has 256K context length, multi-modal and tool calling capabilities, and can be used out of the box with transformers, llama.cpp and MLX. Demis Hassabis confirmed it can run locally on laptops with 16GB VRAM and has over 150 million downloads. — via 1 2
- OpenAI launched two major updates: the ChatGPT memory system that enables context retention across conversations, and new features for GPT-Rosalind, its enterprise life sciences model, combining GPT-5.5 agent coding and tool use with drug discovery and analysis capabilities. — via 1 2
二、Hugging Face Ecosystem Updates
- Hugging Face launched LeLab, a GUI for LeRobot that allows users to complete robot learning workflows without command line tools, with zero-terminal setup, one-click data collection and GPU training to lower the barrier to AI robotics entry. — via 1
- The platform updated TRL to support fine-tuning models directly on agent trace data, with compatibility for Claude Code, Codex, OpenClaw, Pi and other trace data formats, and announced upcoming full support paired with datasets v5 for training custom agents with local, synthetic or community trace data. — via 1 2
- Hugging Face launched the SynthTraces project, which generates over 2000 coded agent session trace datasets for training and optimizing large models, with code and data published on the Hugging Face Hub. The platform also recommended storing all agent trace data in private Hugging Face spaces for historical retention, analysis and model optimization. — via 1 2
三、Agent & Developer Tooling Updates
- Perplexity launched Perplexity Computer for enterprises, which integrates over 400 tools including Intuit QuickBooks, Vercel, Shopify and Canva to support end-to-end business workflows, and announced that Perplexity Personal Computer will launch on Windows for local execution, starting with paid Max and Enterprise Max subscribers. The company also partnered with the US Small Business Administration to offer $25 million in Computer credits for 100,000 eligible small businesses. — via 1 2
- Runway released Aleph 2.0, a video editing tool that enables precise content modification while preserving the rest of the frame, available in the new Edit Studio. — via 1
- Andrew Ng launched a short course on efficient LLM deployment in partnership with Red Hat, covering low-latency serving, model quantization, vLLM deployment and benchmarking to balance speed, cost and accuracy. — via 1
四、Industry Observations & Research
- Ethan Mollick shared data showing Claude Mythos achieved 3-4 hour 80% METR task duration in May, ahead of expert predictions, and noted that 80% of code merged into Anthropic's codebase as of May 2026 was written by Claude, with accelerating adoption trends. — via 1 2
- Yann LeCun highlighted visual state tracking as a major challenge for computer vision in the coming years, and criticized a proposal to tie federal research grants to political policy priorities rather than project merit. — via 1 2
