Vercel's AI Gateway now offers NVIDIA's Nemotron 3.5 Lightning at no cost, served by NVIDIA. The open-weight reasoning model has 30 billion parameters with 3 billion active per token, a hybrid Mamba-2 and mixture-of-experts design and a 1M-token context window; the provider may keep prompts to improve the model.
Apple researchers present what they describe as the first masked diffusion models pretrained from scratch on text, image-text and audio-text data. The paper studies multimodal scaling laws, data mixing ratios, noise schedules and batch-size effects, and proposes default settings for inference.
Simon Willison's llm-gemini plugin now supports Gemini 4 Argon, and his SVG tests at three reasoning levels came out well. He notes the launch price of $2 per million input tokens and $10 per million output matches GPT-6 Sol and undercuts Anthropic's models, but is set to double on December 1.
Google DeepMind has published a post on how it approached responsibility and safety in developing Gemini 4 Argon.
Liquid AI presents LFM2.5-24B-A2B on the Hugging Face blog as a model for running fast, private AI agents directly on the user's device.
n8n says it is among the first companies to integrate Google's Gemini 4 Argon. Its post explains how to use the model in existing n8n workflows and build new ones, and reproduces the benchmarks Google published.