NVIDIA open-sources Nemotron 3.5 Super VL (120B total, 12B active) with 24GB and 4GB Bolt edge models

NVIDIA opened up Nemotron 3.5 Super VL on October 15, 2026 — a 120-billion-parameter mixture-of-experts model that activates only 12 billion per token. The same post shipped two smaller edge siblings, Bolt 74B-A9B 24GB and a 4GB Bolt E4B due October 23. Super VL is the multimodal flagship of the Nemotron 3.5 family, not the August Lightning release and not the hard-reasoning Ultra. The Hugging Face card states up to 1M tokens of context and a licence called OpenMDW-1.1 that covers commercial use, with base weights, post-trained weights, NVFP4 checkpoints, datasets, training recipes and the reinforcement-learning environments all published.

Key points

• Super VL is a 120B-total/12B-active MoE with 512 routed experts per MoE layer, a Mamba-2-Transformer hybrid backbone and a C-RADIO v4-H vision encoder, taking text, images and short video as input (HF model card, October 15, 2026).
• Bolt 74B-A9B 24GB stays at 74B total and 9B active but lands near 20.5 GiB through native 2-bit quantization, hitting 83.99 TPS on AIPerf SPEED-Bench on RTX Spark (NVIDIA Research, October 15, 2026).
• Training-scale figures disagree across NVIDIA’s own pages: the official blog says roughly 26 trillion tokens across pre-training and continued pre-training, while the HF card says the foundation corpus exceeds 20 trillion, with text alone above 10 trillion.

Background

Nemotron 3.5 is a family rather than a single launch. The Decoder reads the lineup as Lightning for throughput (a 30B-A3B from August), Ultra for hard reasoning, and now Super VL for multimodal enterprise agents. Super VL is distilled from Nemotron 3 Ultra and was built with the Nemotron Coalition — Mistral AI, Thinking Machines Lab, Perplexity and Black Forest Labs. NVIDIA’s own framing puts the weight on enterprises post-training against their own data on their own infrastructure, away from a provider’s terms of service.

The hardware neighbours matter here. NVIDIA’s blog puts RTX Spark laptops on sale October 16, and The Decoder notes Bolt also fits DGX Spark with up to 128GB of unified memory; this site’s October 2 entry on DGX Spark’s 64GB edition at $4,999 priced the 128GB version at $6,950, and the Surface Ultra launch put RTX Spark laptops at $2,599. A 24GB ceiling lands on a very specific set of machines.

The facts

Every headline benchmark below is NVIDIA’s own, with no independent re-measurement:

Item	Value	Source
Total / active params	120B / 12B	HF model card
Architecture	MoE, 512 routed experts, Mamba-2-Transformer hybrid, C-RADIO vision encoder	HF card, NVIDIA Research
Context	Up to 1M tokens	HF model card
Licence	OpenMDW-1.1	HF model card
SWE-bench Verified	71.07 (NVFP4) / 71.10 (BF16)	HF model card
CharXiv reasoning	79.88 / 80.82	HF model card
GPQA-Diamond	89.90 / 89.52	HF model card
Hermes Bench	63.44%	NVIDIA blog, via IT之家
LangChain behaviour eval	87.05%	NVIDIA blog, via IT之家
Agent throughput, single GPU	Up to 5× comparable open models	NVIDIA citing Artificial Analysis
Where to get it	Hugging Face, OpenRouter, ModelScope, NIM	NVIDIA blog

The training recipe came out with it. NVIDIA says about 26 trillion tokens of pre-training and continued pre-training plus roughly 3.8 million post-training samples, with reinforcement learning across dozens of environments; the HF card breaks the corpus down further — Nemotron-CC-v2/v2.1 at 9.1 trillion tokens and 1.7 trillion tokens of code, and it states the foundation corpus covers English, 19 other spoken languages and 43 programming languages (the card’s own wording, not an Indic-specific count). A separate 100B-token Indic continued-pretraining stage appears in the card, and its dataset table lists ai4bharat/sangraha as an English text source at 109,263,115 indexed source sequences, 77,231,480 of them verified. The two headline numbers do not reconcile (above 20 trillion on the card versus about 26 trillion in the post), so this article holds both.

Not a single customer percentage appears in NVIDIA’s blog, which lists use cases only — Scale AI for Skydance’s vendor-credit finance agent and EY’s M&A due-diligence agent, AfterQuery for Hebbia’s credit-agreement analysis, AI21 and Alice for cybersecurity attack resilience, Ramp for financial document intelligence, Lila Sciences, R1 for medical claims, TrainLoop with dermatology data from Nolla. The numbers come from IT之家 and The Decoder, and the two do not agree. IT之家 has Scale AI and Skydance expecting inference cost down 97.5% and latency down 80% (roughly one-fortieth and one-fifth), Ramp expecting up to 50% higher throughput and 90% lower latency, and AI21 and Alice seeing a 40% accuracy gain on threat intelligence and vulnerability analysis. The Decoder writes the first pair as costs near one-fortieth and response time near one-fifth, gives Ramp ten-times-faster responses, and attributes the 40% to NVIDIA. No source publishes a methodology for any of them.

The edge trade-offs

Bolt 74B-A9B 24GB is the piece closest to a developer’s desk. NVIDIA Research’s page explains that Star Elastic — a router-guided architecture search and compression framework — squeezed Super VL’s 12B active parameters down to 9B: a budget-conditioned router picks routed-expert counts and MoE widths per layer, with top-K fixed at 9 against the parent’s 22 and the search target set at a footprint under 24GB. The claim is that compressing beats training a same-size model from scratch.

Quantization is mixed by layer rather than uniform: routed and shared experts go to IQ2_XXS at 2.0625 bits per element with quantization-aware distillation to recover accuracy, Mamba layers stay in NVFP4, and embeddings, attention and lm_head stay at Q8_0. Weights only, no activation quantization, on the argument that inference on RTX/DGX Spark is mostly weight-bandwidth bound. The card puts the footprint below 24GB at batch size 1 and 128K context, across two GGUF files totalling about 20.5 GiB.

The scores are self-reported, and the table already contains a model that edges past it: PinchBench 87.82 against Qwen3.6-35B-A3B’s 88.07, while BrowseComp at 66.75 leads the rest of that comparison. Speed is 83.99 TPS against a design target of at least 80 TPS within 24GB. Bolt itself was trained on 80 billion tokens, against the parent’s 20 trillion-plus. The 4GB Bolt E4B is dated October 23 by both NVIDIA and IT之家, with no parameter count from either — that gap stays unfilled here.

What others say

The Decoder’s headline is the counter-argument: open models sell NVIDIA hardware. It argues Nemotron is tuned throughout for NVIDIA silicon, so anyone wanting the most out of it ends up buying NVIDIA cards, and calls the strategy self-serving because more open-model deployment drives more compute demand. It also flags the 24GB requirement as a real barrier at current GPU prices, leaving local agents to developers and workstation owners for now, and notes that Super VL still trails closed frontier models by a wide margin on the overall Artificial Analysis ranking.

ZDNET is more cautious on what the numbers prove: Artificial Analysis’s evidence demonstrates efficiency gains for now, and the base model does not beat proprietary frontier models on its own. The example ZDNET uses is NVIDIA’s own anonymous post-training case — a 44% gain on medical insurance claims that put the model far above Claude Opus 5 — and it takes that as proof the improvement comes from post-training rather than from the shipped checkpoint. ZDNET names no customer for that claims case; NVIDIA’s blog puts R1 on medical claims and revenue-cycle management, while AI21 and Alice sit on cybersecurity attack resilience there. The 40% figure for AI21 and Alice comes from IT之家, so the two should not be merged.

The Prime Intellect post is the only third-party-training writeup with real detail, and it is a collaboration rather than an independent reproduction: the 2,500 enterprise agent tasks were built jointly by Prime Intellect and the NVIDIA Nemotron team, and NVIDIA ran the training itself on 128 GB200 GPUs using NeMo RL (32 for training, 96 for rollout generation) — GRPO for 100 steps over roughly 11 hours and about 1,450 GPU-hours. AutomationBench-AA on the public split went from 34.0% to 57.7%, a 23.7-point gain, with roughly half of it inside the first 40 steps. The authors state plainly that the run is short and the dataset is small by pretraining standards, and that data quality rather than step count is the bottleneck. They also disclose that about 15% of generated tasks were gameable before dedicated checks were added, and that an audit found three leaks — task data readable from the shell, grader source visible inside the sandbox, and policy text resembling grader instructions.

Our take

What makes Super VL worth a look is not the decimal places on a leaderboard; it is that weights, datasets, recipes and RL environments arrived together, so an enterprise planning to post-train has something concrete to start from instead of guesswork. Bolt is the same idea compressed far enough to sit resident on one consumer card. The honest cost is that every number that matters is self-reported, and the one outside attempt with real detail was a joint effort.

If you have 24GB of memory to spare, run Bolt first and decide about the parent model after. If your task is narrow, 12B active parameters plus the open environments will probably get you further than the benchmark table.

How to try it

Super VL weights live at nvidia/NVIDIA-Nemotron-3.5-Super-VL-120B-A12B-NVFP4 on Hugging Face, and Bolt 24GB at nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF; both carry OpenMDW-1.1. Bolt runs through llama.cpp, with a recommended temperature of 1.0 and top_p of 0.95; the 32GB recipe gives a 262,144-token context with 7 draft tokens of speculative decoding, and the 28GB recipe gives 65,536 tokens with a quantized KV cache. Ollama, LM Studio and vLLM are on the supported list too. If you would rather not download anything, Super VL is already on OpenRouter and NIM, with NIM refreshing every two weeks for security and performance fixes.