NVIDIA opens Nemotron 3.5 Super VL: a 120B agent model for one GPU
A 1M-context multimodal MoE, open under OpenMDW, plus a 24GB Bolt edge sibling, live hosted endpoints, and the recipes to tune it yourself.
Copy markdown
120B total, 12B active — and open
Nemotron 3.5 Super VL is a hybrid Mamba-2/MoE vision-language model: 120B parameters, 12B active per token, up to 1M context, and text + image + video in. NVIDIA reports 71.1 on SWE-bench Verified, 77.5 on BrowseComp, and 52.3 on OSWorld-Verified — and the 4-bit NVFP4 build runs on a single H200, GB200, or DGX Spark. It's released under the permissive OpenMDW-1.1 license.
Bolt: a 24GB coding agent for your laptop
The Bolt 74B-A9B sibling uses native 2-bit quantization to fit in 24GB and runs on an RTX laptop or DGX Spark via Ollama, LM Studio, llama.cpp, or vLLM — NVIDIA pitches it for local multi-step coding, tool calls, and web search. A tiny 4GB Bolt E4B for voice assistants and on-device agents ships Oct 23.
Weights, endpoints, and recipes to retrain
No GPU to spare? Call it today on CoreWeave (~$0.15/$0.75 per 1M tokens), Together, DeepInfra, or OpenRouter, or pull BF16/NVFP4 weights from Hugging Face. NVIDIA also opened the datasets, training recipes, and RL environments in NeMo Gym, so you can post-train or distill your own variant.
Reality check: ~90% of work has no benchmark
Mercor's new Benchmark Map cross-walks 29 agent/LLM benchmarks (924 tasks) to O*NET and finds coverage for just 208 occupations — roughly 90% of US working hours sit in tasks nothing measures. A useful reminder when you're picking a model: a leaderboard win rarely maps cleanly to your actual use case.