NVIDIA Nemotron 3.5 Super VL: the open 120B VLM (12B active) targeting multimodal agents — and the Bolt 74B

NVIDIA Nemotron 3.5 Super VL: the open 120B VLM (12B active) aiming at multimodal agents — and the Bolt 74B

🔎 NVIDIA opens up its agentic VLM — and embraces specialization

On October 15, 2026, NVIDIA released Nemotron 3.5 Super VL, an open weight multimodal model with 120 billion parameters, of which only 12 billion are activated per token. It doesn't arrive alone: two compact models, Nemotron 3.5 Bolt 74B-A9B and Nemotron 3.5 Bolt E4B, complete the series the same day.

The positioning is deliberately narrow. In the official announcement, NVIDIA targets "specialized always-on agents": agents that run continuously and chain together multimodal tasks. Not yet another generalist chatbot — a task engine.

The independent signal came the same day: Artificial Analysis ranks Super VL as the best American open weight model, with 26 on its Intelligence Index. Following the Nemotron 3 Ultra 550B of the previous generation (our analysis), NVIDIA is changing doctrine: fewer active parameters, more throughput, and variants that fit on a single RTX card.

The Essentials

• One family, three models: Super VL 120B-A12B (open weights VLM), Bolt 74B-A9B (24 GB, native 2-bit quantization, available now) and Bolt E4B (4 GB, available October 23, 2026).
• Hybrid architecture: MoE Mamba-2–Transformer + C-RADIO vision encoder, 512 routed experts per layer, context up to 1M tokens, text/image/video inputs (MP4 clips of 2 minutes maximum).
• Independent benchmarks: 26 on Artificial Analysis' Intelligence Index, best American open weight. Bolt 74B-A9B: 21.4, far ahead of the median of 8 for open weights of similar size.
• Vendor claims: up to 5x single-GPU throughput versus open models in the category, GPQA-Diamond 89.90 in NVFP4 — NVIDIA figures, not third-party measurements.
• Availability: Hugging Face, Build.NVIDIA.com, OpenRouter, ModelScope, NIM, and the usual runtimes (vLLM, SGLang, TensorRT-LLM, llama.cpp, Ollama, LM Studio, Unsloth).
• The limitation: open weights under the OpenMDW-1.1 license, but the SFT/RL training data is not published.

Recommended tools

Since Super VL and the Bolt models are open weights, the execution ecosystem matters more than the interface. Here are the options that cover the full cycle, from testing to deployment (prices as of October 2026).

To choose based on your hardware, our guides installing an LLM locally and best local LLMs remain the simplest entry point.

Nemotron 3.5 Super VL: what exactly is it?

It's an open-weight VLM (vision-language model) with 120 billion total parameters, 12 billion of which are activated per token, released under the OpenMDW-1.1 license on October 15, 2026. It accepts text, images, and video as input, and generates text — with a reasoning mode that can be toggled on or off.

The architecture in three minutes

The Hugging Face model card for the BF16 model describes a hybrid Mamba-2–Transformer MoE architecture, paired with a C-RADIO vision encoder. Concretely: part of the network handles long sequences at linear cost, the other retains standard attention, and only 12B parameters activate per token.

The MoE has 512 routed experts per layer. That's what makes possible the trade-off claimed by NVIDIA: the capability of a large model, the running cost of a mid-sized one. The context window goes up to 1M tokens, with native multi-token prediction (MTP).

On the multimodal side, inputs cover text, images, and video — MP4 clips up to 2 minutes long. Reasoning can be switched on or off via the enable_thinking parameter, useful for saving tokens on simple agent tasks.

What it changes on the hardware side

The BF16 variant requires 4× H100 or 2× GB200. The NVFP4 variant, obtained through quantization-aware distillation, runs on a single GPU: 1× DGX Spark, 1× H200, 2× H100, or 1× GB200. That's the entire difference between a cluster and a single machine.

Three models, three targets: Super VL, Bolt 74B and Bolt E4B

Three models were announced on October 15, 2026, each calibrated to a different memory footprint. As Technology.org summarizes, it's an open multimodal with 120B/12B active parameters, accompanied by Bolt models designed for local agents.

The Bolt 74B-A9B is distilled from Super VL, with native 2-bit quantization designed to fit within 24 GB — the territory of consumer RTX cards and the DGX Spark. According to MarkTechPost, NVIDIA ranks it first among open models on PinchBench and second on BrowseComp.

On Artificial Analysis's page, it shows 21 on the Intelligence Index — "well above average" for open weights of similar size, whose median is 8. AA's article places it 4th among models under 100B and 1st among American models in this class, with a score of 21.4.

The Bolt E4B, meanwhile, arrives on October 23. It's the "runs everywhere" variant aimed at embedded agents and machines without a dedicated GPU. NVIDIA is clearly building a funnel: a flagship to establish credibility, compacts to spread adoption.

Benchmarks: what's measured, what's claimed

Two families of numbers need to be distinguished: independent measurements from Artificial Analysis, and the benchmarks published by NVIDIA. Both are useful, but they shouldn't be read the same way.

The independent measurements

Artificial Analysis gives Super VL a score of 26 on its Intelligence Index and declares it a "leading US open weight model". That's the best American open weight score to date — and the "US" qualifier isn't decorative, we'll come back to that below.

For the Bolt 74B-A9B: 21.4, 4th among models under 100B and 1st in the United States in that class. The median of comparable open weights is 8. The jump is real, not cosmetic.

The NVIDIA numbers, to be taken as such

The vendor benchmarks, relayed by MarkTechPost, give a GPQA-Diamond of 89.90 in NVFP4 (best claimed score) and a V* of 85.6 in BF16 versus 85.95 in NVFP4. These are NVIDIA numbers, measured on its own optimizations — to be considered as such.

The most honest number in the series is also the least flattering: 17.78 on tau3-bench banking, the worst claimed score of the lot. Tool-using agentic tasks in a banking environment remain hard for everyone, Super VL included. Keep that number in mind before promising an autonomous agent to a client.

A useful clarification: Artificial Analysis also ranks Super VL 4th among "under 100B" models — presumably in active parameters, since the model only activates 12B per token. And the "leading US open weight" title implies that non-American open weights still lead overall. In the self-hostable model space, competitors like Kimi K2.6 from Moonshot AI or GLM-5 from Z.AI occupy that niche — that's where the real open weights battle is being fought, while the top closed models (GPT-5.5, Gemini 3 Pro Deep Think, Claude Opus 4.7) still dominate the general agentic rankings.

NVIDIA's thesis: specialized agents, not chatbots

NVIDIA isn't selling a chat model; it's selling agent infrastructure. All the communication around Super VL talks about task completion, tool-calling, and "always-on" workloads — never about conversation.

The evidence presented is operational. Three partners — Scale AI, Ramp, and AI21 — post-trained the model and report, according to NVIDIA's technical blog, up to 40x reduction in inference cost, 10x lower latency, and +40% accuracy on their use cases. These are partner figures, so they need context — but they describe a model designed for cost per task, not cost per token.

On the serving side, the announced Dynamo recipes are telling: on 16 B200 GPUs at 50 tokens/s, 8x more users served and first-token latency reduced by 85%. The optimizations cited — LiLiCorr speculative drafting (up to 2.3x), NVFP4, multi-token prediction, context reuse — all target the same goal: chaining agentic loop iterations as fast as possible.

NVIDIA wraps it all up in a "Nemotron Coalition": accuracy comparable to its largest reasoning model for about one-fifth of the parameters, and "5x faster than the next best" in agentic work — always to be read as a vendor announcement.

If you're building agents, two resources frame the topic: choosing the LLM for AI agents and the comparison of the best autonomous agents. And the open source ecosystem isn't sleeping: OpenSeeker-v2 shows that pressure on industrial search agents is also coming from open source.

Running Super VL and the Bolt models: the deployment matrix

It all comes down to one question: how many GB of VRAM do you have? The official matrix leaves no room for ambiguity.

The announced runtimes cover the whole spectrum: vLLM, SGLang and TensorRT-LLM for serious serving, llama.cpp, Ollama and LM Studio for local use, Unsloth for fine-tuning. Our guide to installing a local LLM walks you through getting started, and our feature on AI agents with Ollama shows how to orchestrate these models as agents.

Two GPU-free options are worth a look. First, APIs: OpenRouter and NIM distribute Super VL, and if the budget is zero, our directory FreeLLM API — 34 providers and 635 free endpoints lists what can be tested for free. Then, hosting the orchestration: running your agent scripts on a VPS (for example with Hostinger) while calling the model via API remains the most economical setup to get started.

Open weights, but not open data: what NVIDIA publishes — and what it doesn't

The weights are open, the training recipe is partially documented, but not everything is published. This is the distinction you shouldn't miss.

According to NVIDIA's documentation, Super 3.5 is initialized from Nemotron 3 Super, pre-trained on text and then moved to multimodal SFT, followed by RLVR and multi-teacher distillation (MOPD). The distillation from Nemotron 3 Ultra — the 550B generation whose strategy we had already analyzed — uses a SkewReverse loss targeting coding, a detail that will also interest users of LLMs for code.

One rare element is genuinely open: a dataset of 2.8T tokens covering 22 Indian languages, enriched with an additional 10B Indic tokens. This is a concrete contribution to multilingual open data, beyond the weights.

But the SFT/RL data, for their part, are not published. The OpenMDW-1.1 license covers the weights, not the full recipe. For anyone evaluating a model on reproducibility criteria, this is a nuance that matters — and it still separates open weights from fully open source.

What this release reveals about NVIDIA's strategy

Three readings overlap, and they all hold up.

The first is commercial. NVIDIA sells GPUs, and open models optimized for NVFP4, Dynamo and its own runtimes make its silicon the path of least resistance. An open weight that fully expresses its performance on your hardware isn't a selfless gift — it's an elegant conversion funnel.

The second is doctrinal. After Ultra's 550B, the shift to a 120B/12B-active model and compact Bolts signals the abandonment of the race for size in favor of the race for cost per task. This is the pivot for the entire industry, but NVIDIA executes it with an advantage: it controls the hardware stack all the way down to speculative decoding.

The third is geopolitical. Artificial Analysis' "leading US open weight" label is exactly the narrative NVIDIA needed in the face of the non-US open weights that dominate the overall rankings. A US champion, even if not first globally, is enough to occupy the narrative space. The release of the agent safety platform OpenShell and Sentry at around the same time completes the picture: NVIDIA wants to supply the entire agentic stack, from model to guardrail.

My take: Super VL is a very good American open weight, and the Bolt 74B-A9B is probably the most interesting release of the batch for anyone wanting local agents. But the real innovation isn't the score — it's the packaging: a VLM, two quantized variants, benchmarked serving recipes, and partners attesting to cost per task. It's an offer, not just a checkpoint.

❌ Common Mistakes

Mistake 1: reading "leading US open weight" as "best open weight"

Artificial Analysis's title is precise: Super VL is the best American open weight, not the best open weight overall. Non-American self-hosted models still rank higher globally. Always verify the exact scope of a superlative before citing it.

Mistake 2: confusing vendor benchmarks with independent measurements

The 89.90 GPQA-Diamond (NVFP4) is an NVIDIA figure, obtained on its own optimizations. The only third-party measurements cited here come from Artificial Analysis: 26 for Super VL, 21.4 for the Bolt. Systematically cross-check both before drawing conclusions.

Mistake 3: underestimating the required VRAM

The BF16 120B requires 4× H100 or 2× GB200 — it is not a model "for your PC". Only the NVFP4 variant (1× H200 or DGX Spark) and the Bolts (24 GB and 4 GB) target accessible hardware. Choose the variant before choosing the dream.

Mistake 4: believing that "open weights" means "fully open source"

The weights are under OpenMDW-1.1, but the SFT/RL data is not published and the recipe remains only partially documented, without being reproducible. If reproducibility is a criterion for you, this is not a fully open model.

❓ Frequently Asked Questions

Is Nemotron 3.5 Super VL free?

The weights are freely downloadable from Hugging Face under the OpenMDW-1.1 license, in both BF16 and NVFP4. The cost comes from compute: BF16 requires datacenter hardware, otherwise you have to go through a paid API (OpenRouter, NIM). The model is free, inference is not.

Can you run it locally on a PC?

Not the full Super VL: BF16 requires 4× H100. In NVFP4, a single high-end GPU is enough (H200 or DGX Spark). For a regular PC, the Bolt models are made for that: the 74B-A9B fits in 24 GB of VRAM, and the Bolt E4B, expected on October 23, fits in 4 GB.

What's the difference between Super VL and Bolt 74B-A9B?

Super VL is the flagship VLM: 120B parameters, 12B active, multimodal, 1M context. The Bolt 74B-A9B is distilled from it, with native 2-bit quantization to fit in 24 GB, at the cost of lower accuracy (21.4 vs 26 on the Intelligence Index). One targets the datacenter, the other your local agents.

Does Super VL really process video?

Yes. The model accepts text, image, and video inputs — MP4 clips up to 2 minutes — via its C-RADIO vision encoder. It's a full-fledged VLM, not an LLM with a grafted-on image module, and that's a central argument from NVIDIA for agents that observe continuous streams.

Where can you download the weights?

On Hugging Face (BF16 and NVFP4), on Build.NVIDIA.com with the NIM API in early access, and on ModelScope. OpenRouter also distributes the model via API. The Bolt models follow the same channels, with the Bolt E4B available from October 23, 2026.

How does it compare to other open weights?

Artificial Analysis ranks it as the best American open weight with 26 on its Intelligence Index, and 4th among models under 100B (presumably in active parameters). The Bolt 74B-A9B scores 21.4 against a median of 8 for its class. Overall, non-American open weights remain ahead — the "US" in the title matters.

✅ Conclusion

With Nemotron 3.5 Super VL and its Bolt variants, NVIDIA is no longer chasing the title of biggest model, but rather the best cost-per-task ratio for agents — and on October 15, 2026, all the boxes are checked: open weights, multimodal, 1M context, and variants ranging from 4 GB to 2× GB200. If you're building agents, try the Bolt 74B-A9B locally and compare it against your current stack; our comparison of the best LLMs of October 2026 will help you situate it within the full landscape.