Most open models fall into two camps. Either they are giants that only a data center can run, or they are small enough for a laptop but too weak to trust with real work. Nvidia’s new Nemotron 3.5 family tries to cover both ends at once. It pairs a 120 billion parameter model for companies with a pair of compact models built to live on a desktop GPU, and it releases the weights for all of it.
The short version
- Nemotron 3.5 Super VL has 120B total parameters, only 12B active per token, and reads text, images, documents and video
- Nemotron 3.5 Bolt 74B-A9B is squeezed into 24GB of memory with 2-bit quantization and is available now
- Bolt E4B, a 4GB model for voice and physical agents, is due October 23
- Nvidia says Super VL delivers up to 5x higher agentic throughput on a single GPU than open models of similar size
- The catch: most speed and quality numbers are Nvidia’s own, and independent tests are only starting to appear
What Nvidia actually released
Nvidia announced the lineup on October 15, and TechCrunch covered it on October 19. The family has three parts, and they are aimed at very different buyers.
| Model | Size | Built for | Status |
|---|---|---|---|
| Super VL | 120B total, 12B active | Always-on enterprise agents | Available now |
| Bolt 74B-A9B | 74B total, 9B active, fits in 24GB | Local agents on RTX and DGX Spark machines | Available now |
| Bolt E4B | 4GB | Voice assistants and physical agents | October 23 |
All three use a mixture-of-experts design. That means the model holds a large pool of specialist parameters but only wakes a small slice of them for each token. It is the reason a 120B model can behave like a 12B model when it comes to speed and cost, while still drawing on a much larger body of learned knowledge.
Total size versus the part that does the work
The number that matters for running costs is the active parameter count, not the headline size. Here is how the new models compare, with Meta’s Muse Spark Open added for context.
Figures from Nvidia and Meta announcements. Light bars show total parameters, dark bars show active parameters.
We covered the other end of that chart when Meta’s open model that needs a terabyte of RAM arrived last week. Nvidia is making the opposite bet: the useful frontier for open models may be the ones ordinary hardware can actually load.
Super VL: built to run all day
Super VL is the enterprise piece. Nvidia pitches it as a model for agents that run continuously, reading documents, watching video, calling tools and handing results to other systems. It supports a context window of around one million tokens, which matters when an agent has to keep a long task in view.
Nvidia says the model matches its larger Nemotron 3 Ultra on instruction following and tool use, despite being much smaller. Early adopters named in coverage include Skydance, EY, Ramp and Salesforce. The weights ship under the OpenMDW-1.1 license, which allows commercial use, and companies can post-train the model on their own data and run it on their own servers.
Independent numbers are mixed. Artificial Analysis gives Super VL an Intelligence Index score of 26, which it reports as first among open-weight models in the 40B to 150B class. One aggregator lists 71.1 on SWE-bench Verified and 60.7 on Terminal-Bench 2.1, but the same listing shows it trailing some newer coding-focused open models on agent benchmarks. So it looks strongest as a well-rounded, fast, multimodal workhorse, not as the best pure coder.
Running it in production through Nvidia’s own packaging also costs money. One analysis puts Nvidia AI Enterprise at $4,500 per GPU per year, though that figure comes from a third party and not from Nvidia’s announcement, so check it against current pricing.
Bolt: the 74B model that fits on a gaming card
The more surprising release is Bolt 74B-A9B. Seventy-four billion parameters would normally need well over 100GB of memory. Nvidia compressed it to 2-bit precision so it fits in 24GB, the amount of memory on a high-end consumer graphics card. It runs through llama.cpp, vLLM, Ollama, Unsloth and LM Studio, so developers do not need Nvidia’s own stack to try it.
Nvidia says Bolt ranks first among open models in its class on PinchBench and second on BrowseComp, and one report cites about 84 tokens per second on RTX Spark hardware. A separate analysis argues the model decodes more slowly than some rivals such as Qwen and Gemma. Both claims can be true at once, depending on the task and the settings, which is exactly why you should run your own test before committing.
Reality check on 2-bit models
- Extreme compression saves memory but can cost accuracy on long, multi-step tasks
- Benchmark wins on short tests do not guarantee reliability over a full workday
- Always test on your own data, with your own tools, before trusting an agent
Why the local angle matters
Bolt is aimed at RTX Spark and DGX Spark machines, the same hardware family behind the Surface Laptop Ultra that starts at $2,600. Running an agent locally means your files, emails and screenshots never leave the device, and you pay nothing per token. For anyone who has hesitated to give a cloud agent access to private data, that is a real selling point.
It also fits the wider mood in the industry. Companies are starting to worry about what agents can reach and who watches them. Microsoft just argued that AI agents should be treated like insider threats. An open model you host yourself gives security teams something concrete to inspect, log and restrict, which is harder with a closed API.
What to watch next
- October 23: Bolt E4B arrives, which will show how small an agent model can get before it stops being useful.
- Independent testing: Third-party speed and accuracy runs on Bolt will settle the 84 tokens per second debate.
- Enterprise uptake: Whether Super VL moves beyond the early adopters Nvidia has named.
The bottom line
Nemotron 3.5 is less about topping a leaderboard and more about where AI agents will live. Nvidia is betting that the next wave runs on hardware you own, using weights you can read, with training data and recipes published alongside them. If the compression holds up under real work, a 24GB graphics card just became a serious agent machine. If it does not, you will have lost an afternoon of testing, which is a fair price for finding out.
Sources and further reading
- TechCrunch: Nvidia launches new open models to run agents in the enterprise and on PCs
- Nvidia: Open From Data Center to Desktop, Nemotron 3.5 Super VL and Bolt
- Techzine: Nvidia launches Nemotron 3.5 Super VL and Bolt for AI agents
About this article: GeekBlog covers U.S. technology news, AI, phones, smartwatches and gaming. Every story is written and checked under our Editorial Policy. Spotted a mistake or have a story tip? Contact our editors.
