NVIDIA released Nemotron 3.5 Super VL — an open multimodal model that works with text, images, and mp4 videos up to two minutes long and accepts context up to 1 million tokens. Alongside it, compressed Nemotron 3.5 Bolt 74B-A9B was released: thanks to 2-bit quantization of expert weights, it fits in 24 GB of video memory and runs locally, and on October 23 the lineup will be joined by the edge version Bolt E4B at 4 GB. Weights, datasets, and training recipes are published openly, and both models are already available on Hugging Face and OpenRouter.

image
image

What happened

Nemotron 3.5 Super VL combines 120 billion total and 12 billion active parameters: each MoE layer routes 512 experts, the architecture combines Mamba-2 and transformer layers, and visual input is processed by the C-RADIO encoder. The model was trained on approximately 26 trillion tokens, post-training covered about 3.8 million samples and reinforcement learning in dozens of environments, and generation is accelerated by native speculative decoding through multi-token prediction (MTP). According to vendor benchmarks, the model is on par with Qwen3.5-122B-A10B: GPQA-Diamond — 89.5 versus 86.7; SWE-bench Verified — 71.1 versus 71.9; OSWorld-Verified — 51.1 versus 51.5; OCRBench v2 EN — 68.1 versus 68.3. At the same time, Nemotron 3.5 Bolt 74B-A9B 24GB was released with 74 billion total and 9 billion active parameters, 2-bit quantization of expert weights, and a 24 GB VRAM footprint for local deployment on RTX and DGX Spark; the edge version Bolt E4B at 4 GB is announced for October 23.

Context

The significance of the release is broader than a single model: NVIDIA is building out a lineup of open agentic models along a “from data center to desktop” scheme, where server-grade Super VL 120B, local Bolt 74B-A9B, and the upcoming edge version Bolt E4B cover different hardware classes. Super VL's visual capabilities are aimed primarily at documents — text recognition, tables, charts, and PDFs, i.e., material typical of multi-step business processes, while the combination of Mamba-2 linear layers, wide MoE, and long context is designed for working with large documents and videos. The competitive backdrop is sober: newer Qwen3.8-Flash-Next and GLM-5.3-Flash outperform Nemotron 3.5 on most agentic tests, so the release reads not as a flagship claim, but as an attempt to establish an open stack for agentic scenarios.

Why this matters for the industry

For the industry, the key points are deployment options and freedom from closed APIs: under the OpenMDW-1.1 license, weights, datasets, and training recipes are published, so companies can post-train the model on their own data and deploy it on their own infrastructure. According to NVIDIA, NIM microservices and NVIDIA Dynamo serve up to 8x more users while reducing latency by 85%. The document-centric focus of visual tasks lowers the cost of multi-step agents working with PDFs, tables, and charts, while local Bolt covers on-prem and private deployments where the cloud is not acceptable. For startups in vertical document processes, this is a ready-made agent base without external dependency.

Why this matters for users

Super VL is available on Hugging Face and OpenRouter, but it requires server hardware, so the local path is Bolt 74B-A9B on RTX or DGX Spark: agentic tasks involving web, terminal, and code can be run without the cloud. The depth of reasoning is controlled by the chat template — enable_thinking=True/False plus an intermediate mode — turning the balance between speed and answer quality into manual tuning. A practical first step is to download the weights, run a pilot on your own documents, and compare it with your current API provider on price and quality; the model can be fine-tuned for your vertical without coordinating access with the vendor.

What is still unknown / limitations

All quality figures in the sources come from NVIDIA itself — from the official blog and model cards; independent measurements are not yet available, so any comparisons with Qwen and GLM should be read as a guide, not a verdict. Agentic coding comparisons (SWE-bench Pro, Terminal-Bench 2.1) in the materials are given using different harnesses, which further complicates comparability. The claim that Bolt fits in 24 GB of VRAM is a deployment fact, not proof of preserved quality: there is no pairwise comparison of the quantized Bolt with the BF16 version of Super VL in the materials. The described training recipes are stated processes, not published results of controlled experiments, and the speedup through MTP still needs to be confirmed in independent tests.

Sources

Author

Look at AI, editorial team