AI developer tools · What shipped

NVIDIA has released Nemotron 3.5 Lightning, an open 30B mixture-of-experts model with 3B active parameters designed for high-volume execution in long-running AI agents. The model achieves leading accuracy-speed performance on the Artificial Analysis Intelligence Index, completing agentic tasks up to 30% faster than comparable models while maintaining 86% accuracy on PinchBench. Key features include speculative decoding with multi-token prediction, NVFP4 quantization support, and optimization for agent frameworks like OpenClaw and Hermes Agent. NVIDIA is also introducing NeMo Switchyard, a library for intelligent task routing that directs requests to the most capable and efficient model, alongside releases of training recipes, datasets (including Nemotron-RL Agentic Terminal Pivot for coding agents), and customization tools via NeMo Automodel and NeMo RL. The model is available under OpenMDW-1.1 license and supported across industry tools including vLLM, LM Studio, Ollama, and llama.cpp, with deployment options ranging from DGX Spark to data centers.

oMLX 0.5.8.dev3 adds Meta Muse Glimmer 30B VLM support with DFlash speculative decoding achieving 1.4x decode speedup, makes DeepSeek V4 prefill 23-28% faster across context lengths through kernel optimizations and cache block alignment, and fixes Ling-3.0-flash code accuracy (+17 percentage points on HumanEval). The release also reduces long-context memory for GDN models with SSD cache persistence, removes per-token host sync from grammar-constrained decoding, and fixes tool-call parsing edge cases with improved payload boundary detection.

NVIDIA released several AI developer tools and models optimized for local deployment. Cosmos 3 Edge, a 4-billion-parameter open world model for robotics and vision AI, is now available. Poolside AI launched Laguna S 2.1, a 118-billion-parameter open weight agentic coding model that can run locally on a single DGX Spark. Meta released Muse Glimmer, a 30-billion-parameter dense model with 120K+ context window designed for coding and local agentic AI, delivering over 200 tokens per second on RTX 5090. LTX-2.5, an open-world video generation model, provides up to 20% faster performance and 40% memory savings on NVIDIA RTX 6000 PRO. NVIDIA introduced Nemotron 3.5 Lightning, a 30B mixture-of-experts model delivering up to 4x faster token generation compared to similar models, with support from vLLM, Ollama, llama.cpp and LM Studio. DeepSeek refreshed DeepSeek-V4-Flash, a 284-billion-parameter MoE model with 1 million-token context window available in GGUF format. NVIDIA Sync Cluster Assistant now enables developers to automatically cluster multiple DGX Spark systems for distributed inference. NeMo Switchyard, an open source routing library, automatically directs workflow steps to optimal models and reduced benchmark completion costs to one-third of Opus 4.8.