NVIDIA releases open-weight Nemotron 3.5 Super VL and Bolt

On 15 October 2026, NVIDIA published two posts about Nemotron 3.5 Super VL and Nemotron 3.5 Bolt. Michelle Horton’s technical post is timestamped 15:00 UTC and modified the same day at 16:24 UTC. Kari Briski’s company post is timestamped 15:00 UTC and modified at 15:22 UTC. NVIDIA calls Super VL an open model. The card license, below, is why this account says open weights.

Super VL

The technical post says Super VL is a 120-billion-parameter mixture-of-experts model that activates 12 billion parameters per token. It says the model combines reasoning, instruction following, tool use, and text, image, document, and video understanding. The BF16 card matches the 120B total and 12B active counts. It describes a Mamba-2 and Transformer hybrid with a C-RADIO vision encoder, text, image, and video in, and text out, with context up to 1 million tokens. The card’s release line is “Hugging Face - 10/15/2026.”

What Hugging Face shows

The Hub API, read on 15 October, lists three public repositories. None is private, gated, or disabled.

Repository	createdAt (UTC)	lastModified (UTC)
Super VL BF16	2026-10-12 15:32	2026-10-15 17:04
Bolt 74B BF16	2026-10-09 12:45	2026-10-15 15:42
Bolt 74B 24GB	2026-10-07 17:50	2026-10-15 15:42

The blogs and the Super VL card date the public release to 15 October. The repositories themselves were created earlier and updated on the 15th.

Each card’s frontmatter is license: other and license_name: openmdw-1.1, linking the OpenMDW License Agreement, version 1.1. The Super VL and 24GB roots have no LICENSE file. The Bolt BF16 LICENSE URL is an empty file. The agreement text is the GitHub file the BF16 card points at: permission to deal in the model materials, a patent-litigation termination clause, and a warranty disclaimer. OpenMDW-1.1 has no OSI license page that was found on 15 October, and it does not appear on the OSI licenses index opened the same day, so this account says open weights.

The Bolt BF16 card says that checkpoint is for post-training, and that production inference should use the 24GB pack. The Hub API tags that BF16 repository as a finetune of Super VL. The card text describes compression.

Bolt

The company post says Bolt 74B-A9B 24GB is available now, and that a smaller Bolt E4B 4GB follows on 23 October. E4B was not among the three repositories above. The 24GB card lists 74B total parameters and 9B active, context up to 256K tokens, and precision as IQ2, which it calls 2-bit quantization with quantization-aware training, meant to fit in 24GB. The company post says that pack runs on GeForce RTX PCs, DGX Spark, DGX Station, and RTX Spark laptops it dates to 16 October, with support for llama.cpp, Ollama, LM Studio, Unsloth, and vLLM. That list is NVIDIA’s.

Both Bolt cards say the model was compressed from Super VL with Star Elastic, which they call a router-guided compression method, and they link an arXiv PDF for it. The 24GB card says compression ran on the text tower, the vision tower was reattached, and experts were quantized to IQ2-XXS at 2.0625 bits per element. It lists the BF16 Bolt as its base model, relation quantized.

Throughput and partner figures

None of these figures was rerun here.

The technical post says that on a single GPU, Super VL delivers up to 5x higher throughput for agentic tasks than other open models in its size category. One figure caption says the model delivers over 600 output tokens per second on Artificial Analysis Intelligence Index agentic tasks on a Baseten endpoint, and that the chart marks a 5x speedup. The company post says Super VL is 5x faster than the next best. A later technical-post caption says the same throughput and token efficiency deliver 4x faster task completion. The posts do not use one comparison for every 5x.

The technical post’s opening bullets say Scale AI, Ramp, and AI21 post-trained the model, “achieving improvements such as 40x lower inference cost, 10x lower latency, and 40% higher accuracy.” The examples later in the same post are narrower, and they do not all match that bundle:

• Scale AI, for automated vendor credit processing at Skydance: NVIDIA writes that inference cost “will decrease by 40x and latency by 5x.”
• Ramp, converting receipt images to JSON: NVIDIA writes that Ramp “expects up to 50% higher throughput and 10x lower latency.”
• AI21 and Alice, for threat intelligence and vulnerability analysis: NVIDIA writes that accuracy improved by 41%.

The company post names further customizers, including AfterQuery, Mercor, and Perplexity, without those percentages. The Super VL card prints a benchmark table NVIDIA says it collected with NeMo Gym. Those scores are NVIDIA’s. NVIDIA says an EXO Labs evaluation placed Super VL on the Pareto frontier for accuracy and task completion time, next to a DGX Spark chart.

An earlier note on NVIDIA’s own IOI and IMO scores covers different Nemotron systems.

The weights named above are public. The speed, the partner savings, and the 23 October date for Bolt E4B are claims in NVIDIA’s posts.