NVIDIA is one of the most active publishers of open models, and Nemotron 3.5 is now a full family. The newest piece is Nemotron 3.5 Super, detailed in a technical report published Oct 14, joining Lightning (released in August) and the NeMo Switchyard router.
The lineup
| Nemotron 3.5 Super | Nemotron 3.5 Lightning | |
|---|---|---|
| Size | 120B total, 12B active | 30B total, 3B active |
| Inputs | text, images, video | text |
| Context | up to 1M tokens | up to 1M tokens |
| Built for | planning, coding agents, multimodal tool use | fast, high-volume specialized tasks |
| Released | weights on Hugging Face Oct 12; report Oct 14 | August 2026 |
| Checkpoints | Base BF16, BF16, NVFP4 | Base BF16, BF16, NVFP4 |
Both are mixture-of-experts hybrid Mamba-Transformer models. In plain English: only a small slice of the model runs for each token (that's the "active" number), and part of the network uses Mamba layers, which handle long inputs more cheaply than pure attention. Both have built-in multi-token prediction for faster generation through speculative decoding.
"Active parameters" drive speed. "Total parameters" drive memory. For what that means on your hardware, see how much VRAM you need for a local LLM.
What NVIDIA reports for Super
From the technical report page (NVIDIA's own evaluations):
- 51.78% on SWE-bench Pro, beating Nemotron 3 Ultra with about 4.6x fewer active parameters
- Running agents in parallel with self-verification raised Terminal-Bench 2.1 from 62.7% to 71.9% and BrowseComp from 71.5% to 79.8% on NVIDIA's setups
- The NVFP4 checkpoint keeps 99.62% median accuracy versus BF16 across 36 benchmarks
- Up to 4,856 output tokens/s per GB200 GPU, while holding at least 50 tokens/s per user on a 20-turn agent workload
NVIDIA says Super is validated on A100, H100, B200, GB200 and GB300, plus DGX Spark, DGX Station and RTX-class GPUs. At 120B parameters, though, "runs on RTX" means a quantized checkpoint and plenty of memory, not a typical gaming PC. Our GPU rental cost guide helps if you'd rather rent.
Lightning: the small worker
NVIDIA pitches Lightning as the specialist inside a "system of models". A big model plans; Lightning handles targeted jobs like code review, tool use, security alert monitoring or billing questions. NVIDIA says it gives up to 4x faster output and 30% faster agentic task completion than other models in its class, and runs on RTX PCs, DGX Spark, DGX Station and Jetson.
NeMo Switchyard: picking the model per step
Switchyard is an open-source routing library for agents. Instead of sending every request to one expensive model, it routes each step to the most suitable model in your mix of open, proprietary and NVIDIA models. You can tune it for quality, latency or cost. NVIDIA's internal benchmark claims frontier-level accuracy at nearly one-third the cost of using Opus 4.8 alone.
Routing is the same idea as using a cheap model for easy steps in your own app. Pair it with prompt caching for bigger savings.
License and openness
Super's Hugging Face model cards list the OpenMDW-1.1 license and say the model is "ready for commercial or non-commercial use". NVIDIA also publishes training datasets, recipes and evaluation recipes alongside the checkpoints, which is more than most open-weight releases. As always, read the license for the exact checkpoint you deploy.
Who should care
- Agent builders who want an open model to self-host for planning and tool use: try Super (NVFP4).
- Local-first developers with a DGX Spark or a high-memory workstation: Lightning is the practical pick.
- Teams with big API bills: Switchyard-style routing is worth testing whatever models you use.
Comparing open models this month? See Flint vs Ornith and MiniMax M3.1 Flash.
