Nemotron 3.5 Bolt Hermes setup is one line: ollama launch hermes --model nemotron-3.5-bolt. That starts Hermes Agent on NVIDIA's compact local model, which packs 74B total parameters (9B active per token) into about 24GB, so a 24GB-class machine can run a real agent model with no API bill.
📺 Watch: Nemotron 3.5 Super VL: FREE AI for Hermes Agent!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside
NVIDIA released Bolt on 15 October 2026, alongside Nemotron 3.5 Super VL. Super VL is the big one. Bolt is the one you can actually run at home.
I run Hermes daily inside my Agent OS, with a profile per model. A local profile like this is perfect for private, always-on jobs. Here is what Bolt is, the fast way and the manual way to run it in Hermes, and where it fits.
If you want full Hermes courses and the Agent OS zip, the AI Profit Boardroom is $69/mo with daily updates and four coaching calls a week.
What Nemotron 3.5 Bolt is
Bolt is NVIDIA's Nemotron 3.5 Bolt 74B-A9B. It has 74B total parameters with 9B active per token, and NVIDIA uses 2-bit quantization to pack it into about 24GB. It is aimed at running locally for coding, reasoning and tool use.
Ollama describes it as NVIDIA's compact vision-language model for local coding, reasoning, and tool use. On Ollama it shows up as:
- Name: nemotron-3.5-bolt
- Tags: latest, 74b-a9b and 74b-a9b-iq2
- Size: about 25GB
- Context: 256K
- Input: text and image
- Capabilities: vision, tools and thinking
NVIDIA also publishes an official GGUF on Hugging Face as nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-24GB-GGUF, and Unsloth publishes GGUFs too. NVIDIA says Bolt is directly usable from agents such as Hermes and OpenClaw. A smaller Bolt E4B for about 4GB edge hardware was also announced.
Why Bolt fits a 24GB-class machine
The trick is the mix of a mixture-of-experts design and heavy quantization. Only 9B parameters work on each token, and the whole 74B model is squeezed into about 24GB. So a machine with around that much memory can hold it.
Super VL is a different story. It is 120B total with 12B active and needs roughly 70 to 90GB. That is workstation territory.
So the choice is simple. If your machine is in the 24GB class, Bolt is the Nemotron 3.5 model you can run. If you have far more memory, Super VL is an option. My full guide to Nemotron 3.5 Super VL in Hermes Agent covers that one.
📺 Watch: Hermes AI Agent + Nemotron 3 Nano Omni (FREE + OpenSource!)
Bolt vs Super VL
| Model | Size | Memory | Best for |
|---|---|---|---|
| Nemotron 3.5 Bolt | 74B total, 9B active | About 24GB (Ollama download about 25GB) | Local coding, reasoning and tool use on a 24GB-class machine |
| Nemotron 3.5 Super VL | 120B total, 12B active | Roughly 70 to 90GB | Bigger machines that can hold a much larger local model |
Want Super VL on Ollama instead? See Nemotron 3.5 Super VL on Ollama.
📺 Watch: Hermes3D Is INSANE: AI Agents Now Have a 3D Office
The one-line Nemotron 3.5 Bolt Hermes launch
With Ollama installed, this is the fast route:
- Pull and test the model. Run ollama run nemotron-3.5-bolt. The first run downloads about 25GB, so give it time.
- Launch Hermes on it. Run ollama launch hermes --model nemotron-3.5-bolt.
- Give it a real task. Ask Hermes to read a folder and summarise it, or to call one of your tools. That checks tool use, not just chat.
The same launch command also works for openclaw, opencode and claude, so you can try Bolt across agents.
Prefer to run the GGUF directly? Hermes also runs open models on your own machine through its Local Models feature, which downloads and manages llama.cpp for you, picks the right build for your hardware and handles memory. Per the docs, nothing leaves your computer: no account, no API key and no network access after a model is downloaded. Local Models is on by default in the desktop app on macOS, Windows and Linux. The NVIDIA GGUF and the Unsloth GGUFs are the files to look for if you go that way.
Need help picking which model runs which job in your business? Book a free AI strategy session.
The manual route: point Hermes at Ollama
If you want Bolt inside your existing Hermes setup, or in its own profile, connect Hermes to Ollama yourself. Per the Hermes docs, local Ollama is reached through the custom endpoint flow.
- Start Ollama. Run ollama serve. It listens on port 11434.
- Open the model picker. Run hermes model and choose Custom endpoint (self-hosted / VLLM / etc.).
- Enter the URL. Use http://localhost:11434/v1.
- Skip the API key. Ollama does not need one.
- Enter the model name. Type nemotron-3.5-bolt.
You can also set it in config.yaml with the default model, provider set to custom, and base_url set to http://localhost:11434/v1. The docs note that provider: ollama also works under its own name with a base_url like http://127.0.0.1:11434 and an empty api_key, and a bare host and port gets the /v1 suffix added.
Easiest of all: ask your agent. Paste "add a new Hermes profile that uses my local Ollama at localhost:11434 with the model nemotron-3.5-bolt, and set the context to at least 64,000" and let it do the wiring.
Watch the context length
This is the number one thing that trips people up. The docs say Ollama does not use a model's full context window by default. With less than 24GB of VRAM, the default is 4,096 tokens. With 24 to 48GB, it is 32,768.
Hermes needs at least 64,000 tokens of context for agent work with tools, and smaller windows are rejected at startup. So raise it. The docs recommend starting Ollama with OLLAMA_CONTEXT_LENGTH=64000 ollama serve. You can also bake it into a custom model with a Modelfile. You cannot set context length through the OpenAI-compatible API.
Good jobs for Bolt in Hermes
Bolt is strongest on the things it was built for: coding, reasoning and tool use, all on your own machine.
- Private file work. Summaries, notes and drafts from folders you do not want leaving your computer.
- Scheduled jobs. Daily cron tasks that would cost money on a paid API, like inbox digests or report drafts.
- Coding help. Small scripts, fixes and explanations in a repo.
- Image input. Bolt takes text and images, so it can read screenshots you hand it.
- A fallback brain. A local profile that keeps working when a cloud provider is down or rate-limited.
I set this up as its own profile so I can switch between local and cloud models without touching my main config. The Agent OS guide shows how I organise that.
Limits to know
- Speed depends on your hardware. Local means your GPU and memory set the pace. A machine near the 24GB line will feel slower than a hosted API.
- It is not a frontier model. It is a compact local model. For the hardest reasoning, a big hosted model still wins.
- The download is big. About 25GB on Ollama.
- Context needs setting. Raise Ollama's context to at least 64,000 or Hermes will reject it.
Want to compare local options side by side? Read my pick of the best Ollama model for Hermes Agent, and see how models do on real agent tasks in Goldie Bench.
Inside the AI Profit Boardroom ($69/mo) I share my local model profiles and update them as new models like Bolt drop.
FAQ
What is the command to run Nemotron 3.5 Bolt in Hermes?
Run ollama launch hermes --model nemotron-3.5-bolt. Or connect Hermes to Ollama at http://localhost:11434/v1 and pick nemotron-3.5-bolt as the model.
How much memory does Bolt need?
NVIDIA packs it into about 24GB with 2-bit quantization, and the Ollama download is about 25GB. It is built for 24GB-class machines.
Does Bolt support tool calling in Hermes?
Yes. Ollama lists vision, tools and thinking as its capabilities, and NVIDIA says it is directly usable from agents such as Hermes.
Should I use Bolt or Super VL?
Use Bolt on a 24GB-class machine. Super VL is 120B total with 12B active and needs roughly 70 to 90GB, so it only makes sense on far bigger hardware.
Want a plan for running local and cloud models together? Grab a free AI strategy session, or book a free AI strategy session once Bolt is running and we will plan the next step. Nemotron 3.5 Bolt Hermes is the easiest way to get a real local agent going.
