NVIDIA's Nemotron 3.5 Super VL: 120B MoE multimodal model with 5x agentic throughput on one GPU
markjeffrey · x · 2026-10-16
NVIDIA's technical blog details Nemotron 3.5 Super VL, an open, customizable multimodal model built for agentic workflows:
- Architecture: 120B-parameter MoE activating 12B per token; combines reasoning, instruction following, tool use, and text/image/document/video understanding in a single model, eliminating separate VLMs.
- Performance: up to 5x higher throughput for agentic tasks vs. same-size open models on a single GPU.
- Adoption: Scale AI, Ramp, and AI21 have post-trained it for specialized workflows, reporting up to 40x lower inference cost, 10x lower latency, and 40% higher accuracy.
- Efficiency: LiLiCorr speculative decoding, NVFP4 quantization with quantization-aware distillation, multi-token prediction, and context reuse; runs on DGX Spark and other platforms.
- Customization: SFT, LoRA, and RL via NVIDIA NeMo, with open datasets on Hugging Face and a usage cookbook on GitHub.
Separately, Always AI was listed as an official NVIDIA Nemotron inference partner in the blog.
Related event: NVIDIA Open-Sources Nemotron 3.5 Models, Fitting 120B into 24GB(33 posts)→
More from Infra
- Berkeley scholars: concentrated chip supply chain makes a hardwired frontier AI pause feasible — wfithian · 2026-10-16
- Microsoft trained MAI-Base-2 on 16k GB200 GPUs in just 38 days, with 21% MFU — scaling01 · 2026-10-16
- Robotics data tokenization landscape: Khala maps peaq, THOR, BitRobot and Axis — 0xSammy · 2026-10-16
- Signal65 benchmarks: Intel Xeon 6 runs agent workloads 11% faster for less money — ryanshrout · 2026-10-16
- Microsoft's MAI-Base-2: 2.8T-param MoE pretrained on 16k GB200s in just 38 days — scaling01 · 2026-10-16
- Cloudflare and OpenAI pilot: real-time network signals bring unseen pages into ChatGPT search — irvinebroque · 2026-10-16