agentos.guide › blog
Nemotron 3.5 Bolt In Hermes: The Quick Start, And How It Fits Your Agent OS
A Nemotron 3.5 Bolt Hermes setup runs NVIDIA's new Nemotron 3.5 Bolt model, a 74B-parameter vision-language model in a native 2-bit build, as the local brain of Hermes Agent.
NVIDIA's model card says it has 9B active parameters, fits a 24GB memory target, reads text, images and video, and is licensed under OpenMDW-1.1.
Hermes added it to its one-click local runtime in v0.22.3 on 21 October 2026, six days after NVIDIA released it on 15 October 2026.
You pick it from Settings, then Providers, then Local Models, or from hermes model in a terminal.
This page is the quick start, plus how Bolt fits an Agent OS workflow.
If you want help turning AI agents into Google and AI search traffic, book a free SEO strategy session with my team.
Nemotron 3.5 Bolt Hermes at a glance
I checked every fact here on NVIDIA's own pages and the official Hermes docs, catalog and release notes.
| Fact | What the official sources say |
|---|---|
| What it is | It is NVIDIA's vision-language model for running everyday agentic tasks locally. |
| Size | It has 74B total parameters and 9B active parameters. |
| Quantisation | It uses native 2-bit IQ2 quantisation with quantisation-aware training. |
| Memory target | NVIDIA designed it to fit in 24GB of memory. |
| Context | It supports up to 256K tokens. |
| Inputs | It takes text, images and video, and it outputs text. |
| Licence | It is governed by OpenMDW-1.1 and is ready for commercial or non-commercial use. |
| Release date | NVIDIA's model cards and blog give 15 October 2026. |
| In Hermes | Hermes v0.22.3 added it to the local runtime on 21 October 2026. |
What Bolt is, in one minute
NVIDIA's Hugging Face model card calls it NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB.
It was compressed from the bigger Nemotron 3.5 Super VL 120B-A12B model using an NVIDIA method called Star Elastic, then quantised to 2 bits.
It is a Mamba-2 and Transformer hybrid mixture-of-experts model with a C-RADIO v4-H vision encoder.
The card says it is optimised for local agent harnesses and agentic tasks involving tool calling, web browsing and terminal use.
NVIDIA's technical blog says the goal was to fit into 24GB of memory and decode at a speed of at least 80 tokens per second.
One detail I liked is that NVIDIA's card lists synthetic Hermes Agent reasoning traces among its training data.
The six-step quick start
This is the order I'd follow.
| Step | What you do | Where |
|---|---|---|
| 1 | Check your machine against NVIDIA's recipes, which start at an RTX card with 32GB or an RTX Spark with 64GB. | NVIDIA's model card |
| 2 | Update Hermes to v0.22.3 or newer. | hermes update, or the Store for the Windows Store app |
| 3 | Install the llama.cpp runtime. | Settings, then Providers, then Local Models |
| 4 | Download Nemotron 3.5 Bolt from the catalog. | Local Models catalog |
| 5 | Read the memory fit colour, then click Use. | Local Models catalog |
| 6 | Use hermes model then Local models if you work in a terminal. | Any terminal |
The Hermes v0.22.3 notes list "Nemotron 3.5 Bolt and one-slot DFlash drafts in the local runtime", and they say this part of the release is undocumented until v0.23.0.
So I read the catalog file in the Hermes repo at the v0.22.4 tag to see exactly what Hermes downloads.
| Catalog detail | What the Hermes catalog file says |
|---|---|
| Display name | The model shows as Nemotron 3.5 Bolt. |
| Description | Hermes calls it NVIDIA's 74B mixture-of-experts in a 2-bit build that sees images. |
| Files | It pulls an IQ2 model file of about 22 GB and a vision file of about 1.7 GB. |
| Draft model | It adds a LiLiCorr draft model of about 1 GB for DFlash speculative decoding. |
| Recommendation | The entry is set so Hermes doesn't auto-recommend it, so you choose it yourself. |
Plan for roughly 25 GB of downloads in total.
If you want my team to look at how AI agents can bring your business more search traffic, book a free one-to-one SEO strategy session.
🔥 Want the exact Nemotron 3.5 Bolt setup I run inside my Agent OS?
Inside the AI Profit Boardroom you get the installable Agent OS, step-by-step video tutorials, four coaching calls a week and 3,400+ members building real automations.
The hardware check
NVIDIA's GGUF card gives two recipes for llama.cpp.
| Hardware | NVIDIA's llama.cpp recipe |
|---|---|
| RTX Spark 128GB or an RTX card with 48GB+ | The 32GB recipe runs a 262,144-token context with up to 7 DFlash draft tokens. |
| RTX Spark 64GB or an RTX card with 32GB+ | The 28GB recipe runs a 65,536-token context with up to 6 draft tokens and a q8_0 KV cache. |
The card lists NVIDIA Ada Lovelace and Blackwell GPUs, with Windows and Linux as the supported operating systems.
It doesn't mention Macs, so I can't tell you how it behaves on Apple Silicon.
The 24GB in the name is the model's memory target, while NVIDIA's own recipes ask for more room, because context, the vision file and the draft model need memory too.
The Hermes Local Models docs say each catalog row shows green when it fits your GPU, amber when it spills into system RAM and red when it's too big.
How Bolt fits an Agent OS workflow
An Agent OS is a crew of agents sharing one memory, and each agent can run on a different brain.
My own Agent OS runs agents like Hermes, Claude, GLM and Kimi under one shared memory.
Bolt's job in a setup like that is the private local brain, sitting on the one machine in the building with a big NVIDIA card.
| Agent OS job | Brain I'd give it | Why |
|---|---|---|
| Private client documents | I'd use Nemotron 3.5 Bolt on a local NVIDIA box. | The docs say local models send nothing off the machine after download. |
| Screenshot and chart reading | I'd use Nemotron 3.5 Bolt. | NVIDIA's card says it takes images and short videos. |
| Research and browsing | I'd use Nemotron 3.5 Bolt. | NVIDIA lists web browsing as a target task, with a BrowseComp score of 66.75. |
| Long, heavy jobs | I'd use Nemotron 3.5 Super VL through NVIDIA NIM. | NVIDIA lists Super VL at 120B total with up to 1M tokens of context. |
| Fast small sub-tasks | I'd use a smaller catalog model. | The docs say an 8 GB GPU runs the small catalog models comfortably. |
The Hermes desktop app can register several Hermes backends, so the docs say you can use a local runtime, a remote gateway on your LAN, an SSH host or Hermes Cloud side by side.
That means the NVIDIA box can run Bolt while your laptop talks to it through Settings, then Gateways.
The docs say idle local models unload after 15 minutes to free GPU memory and reload on the next message, which suits an always-on brain box.
🔥 Want the exact Nemotron 3.5 Bolt setup I run inside my Agent OS?
Inside the AI Profit Boardroom you get the installable Agent OS, step-by-step video tutorials, four coaching calls a week and 3,400+ members building real automations.
Running Bolt with your own llama-server
Some operators prefer to run the model server themselves.
NVIDIA's card gives full llama.cpp commands, including --spec-type draft-dflash and --spec-draft-model for the LiLiCorr drafter.
The Hermes docs say that if a llama-server is already running on your machine, Hermes detects it and uses it instead of starting its own, and the default probe is port 8080.
If your server uses another port, the docs say you can list it under local_runtime.detect_ports or define it under providers in config.yaml.
If you'd like help planning which machines and brains your business actually needs, grab a free one-to-one strategy session and we'll look at your site, your competitors and the keywords you're closest to winning.
Adding Super VL for the heavy jobs
The bigger brother is Nemotron 3.5 Super VL, which NVIDIA's card lists at 120B total and 12B active parameters, with up to 1M tokens of context.
The Hermes v0.22.1 and v0.22.2 notes added Super 3.5 default parameters, put Nemotron on the NVIDIA NIM default model list and switched to the released name, Nemotron 3.5 Super VL.
The Hermes providers docs say to add an NVIDIA API key from build.nvidia.com to ./.hermes/.env and run this.
hermes chat --provider nvidia --model nvidia/nemotron-3.5-super-vl-120b-a12bFor a local NIM on a DGX Spark, a DGX Station or a local GPU, the docs say to set NVIDIA_BASE_URL=http://localhost:8000/v1 and use NVIDIA_API_KEY=local-no-auth if your endpoint has no authentication.
In an Agent OS, that gives you one model family for both the local brain and the heavy cloud brain.
What NVIDIA says it can do
These are NVIDIA's own numbers from its model card, so treat them as the maker's claims.
NVIDIA's table shows Bolt at 50.98 on Terminal-Bench 2.1 against 44.38 for Qwen3.6-35B-A3B, and 66.75 on BrowseComp against 48.74.
Qwen3.6-35B-A3B edges it on PinchBench, with 88.07 against Bolt's 87.82.
NVIDIA's blog reports 83.99 tokens per second on its coding speed test on an RTX Spark laptop with 128GB, while Qwen3.6 35B-A3B hit 96.63 and Gemma 4 26B-A4B hit 100.56 in the same table.
So Bolt's pitch is more intelligence per GB, not the fastest output.
Limits before you commit
The hardware bar is high, because NVIDIA's recipes start at 32GB RTX cards and 64GB Spark machines.
The Hermes Local Models docs still say Hermes never offers builds below 4-bit, while Bolt is NVIDIA's own 2-bit build, so the docs haven't caught up with the catalog.
NVIDIA's blog lists the checkpoints under the name "Nemotron 3.5 Super VL 74B-A9B", while its model cards call them Bolt, so search for Bolt on Hugging Face.
No Nemotron model is on GoldieBench yet, so I have no score of my own to give you.
I haven't run Bolt myself, and everything here comes from NVIDIA's pages, the Hermes docs, the Hermes catalog file and the official release notes.
I have run Hermes on a free local model, which I showed in my Gemma 4 video using Ollama and a custom endpoint.
Also On Our Network
FAQ
What is Nemotron 3.5 Bolt?
Nemotron 3.5 Bolt 74B-A9B 24GB is an NVIDIA vision-language model built to run everyday agentic tasks locally. NVIDIA's model card says it has 74B total and 9B active parameters, uses native 2-bit quantisation to fit in 24GB of memory, supports up to 256K tokens and is licensed under OpenMDW-1.1.
How do I run Nemotron 3.5 Bolt in Hermes?
Update Hermes to v0.22.3 or newer, then open Settings, Providers, Local Models, install the runtime, download Nemotron 3.5 Bolt and click Use. In a terminal, hermes model then Local models runs the same setup.
What hardware does Nemotron 3.5 Bolt need?
NVIDIA's GGUF card gives llama.cpp recipes for RTX Spark with 64GB or 128GB and RTX graphics cards with 32GB or 48GB and up, on NVIDIA Ada Lovelace or Blackwell GPUs running Windows or Linux.
When was Nemotron 3.5 Bolt released?
NVIDIA's model cards and technical blog give 15 October 2026. Hermes added it to the local runtime in v0.22.3 on 21 October 2026.
How do I use Nemotron 3.5 Super VL in Hermes?
The Hermes docs use the NVIDIA NIM provider. Add an NVIDIA_API_KEY from build.nvidia.com to ~/.hermes/.env and run hermes chat --provider nvidia --model nvidia/nemotron-3.5-super-vl-120b-a12b, or point NVIDIA_BASE_URL at a local NIM endpoint.
Is Nemotron 3.5 Bolt on GoldieBench?
No. No Nemotron model is on the GoldieBench leaderboard yet, so there is no GoldieBench score for Bolt.
Final word on Nemotron 3.5 Bolt in Hermes
Check your card, update Hermes, download Bolt from Local Models and give it the private jobs in your Agent OS.
If you want a clear plan for getting your business found in Google and AI search, book your free strategy session here.
If you want the whole Agent OS with shared memory, skills and agents already wired in, it lives inside the AI Profit Boardroom, with 3,400+ members and four coaching calls a week.
That is the fastest way to put a Nemotron 3.5 Bolt Hermes setup to work.
About Julian
I'm Julian Goldie, an AI entrepreneur, SEO expert and the founder of the AI Profit Boardroom, which has 3,400+ members.
I help business owners scale with AI agents, automation and SEO.
My YouTube channel has 400,000+ subscribers, and I run Goldie Agency, a seven-figure SEO agency.
Related guides on agentos.guide
→ Hermes Agent v0.22: The Threshold Release
→ The Best Local Model For Hermes Agent
Get the whole Agent OS
The installable system behind every guide on this site — dashboard, agents, memory and pipelines — updated as new tools land, with 3,400+ members and four coaching calls a week.
Join the AI Profit Boardroom →📺 Video notes + links to the tools 👉
🎥 Learn how I make these videos 👉