NVIDIA unveils Nemotron 3.5 Lightning agentic models with open weights and 1M-token context

Nemotron 3.5 Lightning is NVIDIA's play for the persistent-agent workload tier, positioned as the fastest open model in its class. The architecture is a hybrid MoE with 30 billion total parameters and only 3 billion active per token, delivering throughput suited to high-volume enterprise automation — personal assistants, financial document processing, and cybersecurity triage — while keeping inference costs manageable.

Crucially, it's part of NVIDIA's open family with open weights, training data and recipes. The full-precision BF16 variant is offered explicitly as a reference for customization: post-training, domain adaptation, and producing quantized versions. The 1-million-token context window targets long-horizon agent tasks. NVIDIA also shipped NeMo AutoModel support to fine-tune Meta's newly released Muse Glimmer, positioning itself as the tooling layer beneath the open-weight wave.

Early enterprise adoption gives the launch credibility: CrowdStrike (security), Harvey (legal), CodeRabbit (code review) and Lila Sciences (scientific reasoning) have already customized the model. Distribution is broad — OpenRouter, build.nvidia.com as a NIM microservice, NeMo Switchyard, and SageMaker JumpStart.

Competitively, Nemotron slots against Qwen, Mistral and Meta's Glimmer in the open-weight agentic tier, but NVIDIA's differentiator is the full stack: silicon, inference tooling, and now competitive open models. Skeptics will note that NVIDIA benefits regardless of which open model wins, since all run on its GPUs — the model release is as much ecosystem strategy as a standalone product. What to watch: independent benchmarks validating the speed and reasoning claims.