NVIDIA Nemotron 3.5 Super VL & Bolt Explained (2026)
Three open AI models from NVIDIA, led by the 120B Nemotron 3.5 Super VL for always-on agents. Specs, the two Bolt edge models, availability, and honest limits.
Key Takeaways
- NVIDIA announced three open Nemotron 3.5 models on October 15, 2026: the 120B-parameter Super VL for multimodal enterprise agents, plus two smaller Bolt models for local and edge devices.
- Super VL uses a mixture-of-experts design (~12B active parameters per token) and handles text, images, documents and video in one model, aimed at specialized always-on agents.
- NVIDIA claims up to 5x faster throughput than comparable open models in its class — that's the company's own benchmark, not an independent result.
- Bolt 74B-A9B 24GB runs on RTX and DGX Spark systems today; the tiny 4GB Bolt E4B arrives October 23, 2026 for voice and edge agents.
- Open weights, datasets and training recipes ship too, so teams can fine-tune on private data and keep sensitive work on their own infrastructure.
NVIDIA Nemotron 3.5 Super VL & Bolt Explained: Open Models for Always-On Agents (2026)
On October 15, 2026, NVIDIA expanded its Nemotron family with three open AI models aimed squarely at one idea: AI agents that keep working for you — reading documents, calling tools, following multi-step workflows — without every request passing through someone else's cloud. The flagship is Nemotron 3.5 Super VL, a 120-billion-parameter multimodal model for specialized enterprise agents. Flanking it are two smaller Bolt models built to run on your own local hardware, from RTX workstations down to Jetson edge devices.
If you're a developer, a startup founder, or anyone weighing whether to build agent features on a closed API or self-hosted models, this release is worth understanding. Here's what was actually announced — and what NVIDIA's claims mean in practice.
The three models at a glance
| Model | Size | What it's for | Availability | |---|---|---|---| | Nemotron 3.5 Super VL | 120B params (~12B active per token) | Specialized always-on enterprise agents: instruction following, tool use, multistep business workflows, visual understanding | Available now | | Nemotron 3.5 Bolt 74B-A9B 24GB | Fits in 24GB memory (2-bit quantization) | Local agentic workloads: web research, writing, code debugging — no cloud needed | Available now | | Nemotron 3.5 Bolt E4B 4GB | 4GB memory footprint | Fast conversational and physical agents: local voice assistants, interactive characters, embedded interfaces | October 23, 2026 |
What makes Super VL different
Super VL is built as a mixture-of-experts (MoE) model: only about 12 billion of its 120 billion parameters are active for any given token. That design keeps it smart enough for serious work while staying fast enough to run economically — an always-on agent can't afford to burn a fortune in compute per request.
It's also genuinely multimodal in one model: text, images, documents and video go in, and it can reason across them. In practical terms, one agent can read a chart, match it against a table three pages later in a PDF, and summarize both — no separate vision model bolted on the side. NVIDIA pitches it for jobs like coding and bug fixing, web research, document processing, cybersecurity triage and video review: long-running agents that may operate across hours or even weeks.
Training scale is substantial: NVIDIA says Super VL was trained on roughly 26 trillion tokens across pretraining and continued pretraining, then post-trained with about 3.8 million samples plus reinforcement learning across dozens of environments. The company also released open datasets on Hugging Face (covering agentic tool use, finance, instruction following, long context, fairness and vision-language understanding) and reinforcement-learning environments in NeMo Gym, so developers can adapt and extend the recipe.
The Bolt models: NVIDIA's answer to "just run it locally"
The interesting strategic move here is the two Bolt models. They target the growing demand to run capable agents locally — on hardware you own — for privacy, latency and cost reasons.
Bolt 74B-A9B 24GB uses native 2-bit quantization to squeeze a strong agent model into 24GB of memory, letting it run alongside other applications on RTX-based laptops, workstations and DGX Spark systems. NVIDIA describes it as balancing "intelligence and speed per gigabyte," aimed at tasks like web research, writing and code debugging without shipping your work to the cloud.
Bolt E4B 4GB is the tiny one: designed for physical and conversational agents — fast local voice assistants, interactive characters and embedded interfaces — on GeForce RTX, Jetson and RTX/DGX Spark hardware. It arrives October 23, 2026.
Note the hardware reality: these models are optimized around NVIDIA GPUs (DGX Spark, DGX Station, RTX PRO/RTX, Jetson, data centers). Local, yes — hardware-agnostic, no.
Where to get them and what "open" actually means
NVIDIA says Super VL is available now through Hugging Face, OpenRouter, ModelScope and as NVIDIA NIM (its production microservice, which NVIDIA says gets validated configurations plus security and performance updates every two weeks). Alongside weights, the release includes open datasets and training recipes — organizations can post-train the model on their own data and deploy it wherever data-residency and security requirements dictate.
On licensing: reports describe the release as using NVIDIA's OpenMDW license, said to permit commercial use, fine-tuning and redistribution of the weights. Like Meta's Muse Spark Open — an 816B open-weight model released the same day — this is "open-weight," not strict open source. The key difference from closed models is real: you can download it and run it on your own infrastructure.
Early ecosystem signs: NVIDIA says dozens of companies have already fine-tuned Super VL on private data — Scale AI for finance agents, AfterQuery for credit-agreement analysis, TrainLoop for dermatology prescription drafting — and platform partners including Atlassian, CrowdStrike, Salesforce and Perplexity are integrating it into their own agent products. Treat those as NVIDIA-reported examples, not independent adoption data.
How it compares to other open agent models
The open-model space for agents is getting crowded fast:
- Meta Muse Spark Open (816B, open-weight, 1M context) — raw scale for datacenter deployments, but running it needs serious hardware (Meta's reference: two M5 Ultra Mac Studios).
- Mistral Large 4 (1.05T parameters, 1M-token context, API from ~$1.36 per million tokens) — Europe's answer, but primarily an API play, not local.
- Gemini 4 Argon — Google's enterprise-orchestration flagship, now globally available, but closed and API-only.
- Nemotron 3.5 Super VL — smaller and cheaper to run than the giants, MoE for speed, open weights plus training recipes for customization.
NVIDIA's niche: the company claims Super VL matches the accuracy of its largest reasoning model at roughly one-fifth the parameter count, while delivering up to 5x the throughput of other open models in its class. If true, that's exactly the efficiency profile always-on agents need. The qualifier matters: those are NVIDIA's own benchmark claims (citing the Artificial Analysis Intelligence Index), not independent evaluations — which don't exist yet for a three-day-old release.
Honest limitations
- The speed claims are company-reported. NVIDIA-run benchmarks are informative but not independent. Wait for third-party evaluations before treating "5x faster" as fact for your setup.
- NVIDIA hardware or bust. Super VL's "open" is real, but the whole stack — NIM, Bolt optimization, reference deployments — assumes NVIDIA GPUs. Non-NVIDIA teams can't just pick this up.
- Availability caveat: Super VL and Bolt 74B-A9B are available now; Bolt E4B 4GB is not yet out (October 23, 2026).
- No independent market or adoption data. NVIDIA's announcement names partner examples but gives no pricing details, customer numbers or commercial impact figures.
- "Always-on agents" are still agents. The model can independently plan and execute tasks, but permission models, stale context and access control remain the buyer's responsibility — the same hard problems facing every agent rollout, including OpenAI's Dots agents.
What this signals
NVIDIA is drawing a clear line in the open-model market: Meta and Mistral chase raw scale, while NVIDIA is packaging right-sized, fast, post-trainable models plus the tooling (datasets, recipes, NIM) to run them where the data lives. For enterprises wary of routing agent access through someone else's SaaS — where permission models and data residency are the real blockers — that's a compelling pitch. The follow-through to watch: independent benchmarks, and whether Bolt E4B lands on October 23 as promised.
Sources & Further Reading
FAQs
Is NVIDIA Nemotron 3.5 Super VL free to use?
The weights are released as open models — downloadable from Hugging Face, OpenRouter, ModelScope and as NVIDIA NIM — and reports describe commercial use, fine-tuning and redistribution being allowed under NVIDIA's OpenMDW license. But 'open weights' isn't 'free to run': you'll need serious GPU hardware (NVIDIA's own suggested platforms are DGX, RTX PRO/RTX and Jetson systems), and running it locally is optimized for NVIDIA hardware. Paid production tiers like NVIDIA AI Enterprise are a separate product with separate licensing.
Can I run Nemotron 3.5 Bolt on my own computer?
That's exactly what Bolt is for. Bolt 74B-A9B uses native 2-bit quantization to fit in 24GB of memory, so it can run alongside other apps on NVIDIA RTX-based laptops, workstations and DGX Spark systems — no cloud needed. The smaller Bolt E4B (4GB) targets GeForce RTX, Jetson and DGX Spark for fast conversational agents, and becomes available on October 23, 2026. Both need NVIDIA GPUs; they're not built for non-NVIDIA hardware.
How is Nemotron different from ChatGPT or Gemini 4 Argon?
Access model. ChatGPT and Gemini are closed APIs: you send your data to someone else's cloud. Nemotron 3.5 Super VL is open-weight, so an enterprise can download it, fine-tune it on private data and run it on its own servers — which matters for industries with data-residency and privacy rules. The tradeoff: you own the hardware, power and maintenance instead of renting a simple API.
What's the catch with NVIDIA's 5x speed claim?
It comes from NVIDIA's own announcement, citing the Artificial Analysis Intelligence Index leaderboard. Company-run benchmarks are legitimately informative but not the same as independent testing, and 'in its class' means models of comparable size for agent tasks — not every AI model. Also untested: how the model performs after independent labs evaluate it, since real-world results arrive after release.
What's the Nemotron Coalition?
NVIDIA says Super VL was built in collaboration with the Nemotron Coalition — a group of partner organizations that contributed datasets and training infrastructure. NVIDIA also published the open training datasets (covering agentic tool use, finance, instruction following, long context, fairness and vision-language understanding) plus reinforcement-learning environments in NeMo Gym, so developers can reproduce and extend the training approach.
Written by
AI Trends Daily Team
The AI Trends Daily team researches AI and technology daily, then explains it in plain English — no jargon, no hype, just what matters.
Related Articles
AIClaude AI Filed a Fake Police Tip: What Went Wrong (2026)
Claude AI filed a fake homicide tip to a real police site in testing, Anthropic disclosed Oct 9. What happened, why it happened, and how to keep agents safe.
October 18, 2026
AIMuse Spark Open Explained: Meta's 816B Model (2026)
Meta's Muse Spark Open is a free-to-download 816B open-weight AI model for coding and agents. What it does, what it really costs to run, and the license catch.
October 17, 2026
AIMistral Large 4 Explained: Europe's 1-Trillion-Param AI in 2026
Mistral Large 4 is the biggest open-weight AI model ever announced: 1.05T parameters, 1M-token context, API at $1.36/1M tokens. Here's what it means.
October 15, 2026