Nemotron 3.5 Super VL: Nvidia's Open Agent Model
Nvidia ships an open 120-billion-parameter multimodal model for round-the-clock agents, but the headline speed numbers assume Nvidia silicon underneath.
In short
With Nemotron 3.5 Super VL, Nvidia is releasing an open multimodal model of 120 billion parameters, 12 billion of them active per token, built for enterprise agents that run nonstop, while the promised speed advantage shows up mainly on the company's own GPUs and racks.
At a glance
- Super VL carries 120 billion parameters, 12 billion active per token, and also ships as a 4-bit NVFP4 build.
- Pre-training used more than 25 trillion tokens; post-training added roughly 4 million examples across about 100 RL environments.
- Bolt 74B-A9B fits in 24GB of VRAM at 2-bit quantization; a 4GB variant called E4B is due later in October.
- Artificial Analysis Intelligence Index v4.3.2: Super VL 26 points, Bolt 21 points, with Claude Opus 5.5 ahead at 58.
- With Nvidia Dynamo on 16 B200 GPUs: concurrent users rise from about 60 to over 500, time-to-first-token 2.75s to 0.4s.
Nvidia has introduced Nemotron 3.5 Super VL, an open multimodal model meant to keep enterprise agents working around the clock: 120 billion parameters in total, 12 billion of them active for any given token. The speed story around it, however, is inseparable from Nvidia's own accelerators.
What is inside Super VL
Super VL is a mixture-of-experts model shipped both in BF16 and in a 4-bit NVFP4 build. Nvidia puts pre-training at more than 25 trillion tokens and post-training at roughly 4 million examples drawn from about 100 reinforcement learning environments. On the vision side the target workloads are charts, OCR, tables, multi-page documents and video. Weights are on Hugging Face, including NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-BF16.
Benchmarks: fast, not frontier
On the Artificial Analysis Intelligence Index v4.3.2 Super VL scores 26 points, well behind the leader Claude Opus 5.5 at 58; the smaller Bolt model lands at 21. For chart reading on CharXiv, Nvidia reports 80.8%, rising to 84% when the model is allowed to run Python, against 63% for the older Nemotron 3 Nano Omni.
On SWE-bench Pro the company claims about 51% accuracy and roughly 10 hours to work through 10,000 tasks on four B200 GPUs, where it says comparable models need 50 to 56 hours. At an input of 100,000 tokens Nvidia cites around 630 tokens per second and up to five times the throughput of open models in the same class. For Indian languages it reports 58 points on Indic FrontierMath, versus 27 for Sarvam-105B and 22 for Gemma-4-31B-it.
Two small models for local machines
Nemotron 3.5 Bolt 74B-A9B has 74 billion parameters with 9 billion active per step and squeezes into 24GB of VRAM through 2-bit quantization. Nvidia names DGX Spark, RTX Spark, GeForce RTX, DGX Station and RTX PRO workstations as targets, with llama.cpp, Ollama, LM Studio and vLLM as runtimes. On GDPval-v2 the 24GB build sits just under 28 points at about 78 tokens per second. A 4GB version, E4B, is expected later in October for voice assistants, interactive characters and on-device interfaces.
The catch: the speed is hardware-shaped
The most striking figures only appear once Nvidia's own serving stack is in play. With the Dynamo inference software the company reports eight times more concurrent users and an 85% cut in time-to-first-token; on 16 B200 GPUs concurrency moves from about 60 to more than 500 and the first token arrives after 0.4s instead of 2.75s. Speculative decoding is said to deliver up to 2.3x faster text generation on DGX Spark. Open weights and proprietary accelerators are doing the work together here, and reaching the quoted numbers means buying the hardware.
Who is already tuning it
Nvidia lists dozens of organizations adapting Super VL: EY for acquisition due diligence, Hebbia for credit agreements, CrowdStrike for network defense, Salesforce for government workflows, plus Perplexity, Dassault Systemes and Mercor. The efficiency figures attached to those projects are customer expectations rather than audited results: Scale AI anticipates costs falling to one-fortieth and response time to one-fifth, Ramp expects ten times faster responses, AI21 a 40 percent gain in accuracy. Nvidia also published training data, among it an Indic set of 2.83 trillion tokens covering 22 Indian languages, built with IIT Madras and AI4Bharat. The Nemotron coalition is reported to include Mistral AI, Thinking Machines Lab, Perplexity and Black Forest Labs.
What we could not confirm
A second intended source for this release could not be opened; the server answered with HTTP 403 on both attempts. This report therefore rests on a single newsroom. That leaves the exact license terms for the weights, the context window and every vendor-supplied benchmark figure independently unverified.
FAQ
How big is Nemotron 3.5 Super VL?
It is a mixture-of-experts model with 120 billion parameters in total and 12 billion active per token, released in BF16 and in a 4-bit NVFP4 build.
Can Nemotron 3.5 Bolt run on a 24GB GPU?
Yes. The 74B-A9B build is quantized to 2-bit for 24GB of VRAM and targets DGX Spark, RTX Spark, GeForce RTX, DGX Station and RTX PRO workstations, with a 4GB E4B variant due later in October.
Where can I download the Nemotron 3.5 weights?
Base, post-trained and 4-bit versions are published on Hugging Face. We could not verify the precise license conditions from the one source that was reachable.
