GoldieBench Blog · 9 min read
Nemotron 3.5 Bolt In Hermes: What It Is, How To Run It, And Where It Sits On GoldieBench
Nemotron 3.5 Bolt in Hermes explained: NVIDIA's 74B 2-bit local model, hardware, licence, setup, and whether any Nemotron is on GoldieBench yet.
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 on Hugging Face 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.
I'm Julian Goldie, and in the video above I walk through the v0.22 release that brought one-click local models to Hermes.
In this post I'll cover what Bolt is, what it needs, how to run it in Hermes and every feature, and then I'll tell you plainly where it stands on GoldieBench.
If you want my team to look at how AI agents and AI search can grow your business, you can book a free SEO strategy session.
| Item | Detail |
|---|---|
| What it is | It is NVIDIA's vision-language model for running everyday agentic tasks locally. |
| Who made it | NVIDIA Corporation made it, and Nous Research added it to Hermes. |
| Size | It has 74B total parameters and 9B active parameters. |
| Quantisation | It uses native 2-bit IQ2 quantisation with quantisation-aware training. |
| Context | It supports up to 256K tokens. |
| 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. |
| On GoldieBench | No Nemotron model is on the board yet. |
What is Nemotron 3.5 Bolt?
NVIDIA's full name for it is NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB.
The model card describes it as a vision-language model designed to run everyday agentic tasks locally, using a native 2-bit quantisation to fit in 24GB of memory and run alongside other applications.
It was compressed from NVIDIA's bigger Nemotron 3.5 Super VL 120B-A12B model using an NVIDIA method called Star Elastic.
NVIDIA's technical blog says the goal was to fit into 24GB and decode at a speed of at least 80 tokens per second.
The card says it is optimised for local agent harnesses and agentic tasks involving tool calling, web browsing and terminal use, which is exactly Hermes Agent's job.
NVIDIA's card also lists synthetic Hermes Agent reasoning traces among its training data.
The full spec sheet
| Spec | What NVIDIA's model card says |
|---|---|
| Architecture | It is a Mamba-2 and Transformer hybrid LatentMoE with a C-RADIO v4-H vision encoder. |
| Precision | The experts use IQ2_XXS at 2.0625 bits per element, the Mamba layers use NVFP4, and the embedding and output layers use Q8_0. |
| Inputs | It takes RGB images, MP4 video up to two minutes, and text. |
| Output | It writes text, including JSON, chain-of-thought reasoning and tool calls. |
| Speculative decoding | It uses DFlash with a separate LiLiCorr drafter that NVIDIA's blog puts at 1.13B parameters. |
| Reasoning | Thinking can be switched on or off through the chat template. |
| Sampling | NVIDIA recommends a temperature of 1.0 and a top_p of 0.95. |
| Runtime | The GGUF version runs on llama.cpp, and a separate 24GB version runs on vLLM. |
| Supported GPUs | The GGUF card lists NVIDIA Ada Lovelace and Blackwell. |
| Operating systems | The GGUF card lists Windows and Linux. |
The problem Bolt solves
Local agents have always had a memory problem.
The models that fit on one machine tend to be small, and small models struggle with long, multi-step agent work full of tool calls.
You can see that on my own board, where the small local models score far below the big cloud ones.
NVIDIA's answer is to take a big model, compress it and quantise it hard instead of training a small one from scratch.
Its blog says larger networks at 1 to 2 bits are more Pareto-optimal than smaller networks at higher bits for the same memory size, and NVIDIA concludes the same from its own tests.
For a Hermes user, that means a much bigger brain on the same GPU, with the privacy of running everything at home.
The Hermes Local Models docs say local models run with no account, no API key and no network access after the download.
What hardware Nemotron 3.5 Bolt needs
NVIDIA's GGUF card gives two llama.cpp recipes.
| Hardware | NVIDIA's recipe |
|---|---|
| RTX Spark 128GB or an RTX card with 48GB+ | The 32GB recipe runs a 262,144-token context with up to 7 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 24GB in the name is the model's memory target, while the recipes need more room for context, the vision file and the draft model.
NVIDIA's card doesn't mention Macs, so I can't tell you how it runs on Apple Silicon.
How to run Nemotron 3.5 Bolt in Hermes
The Hermes v0.22.3 notes list "Nemotron 3.5 Bolt and one-slot DFlash drafts in the local runtime".
The setup follows the Hermes Local Models docs.
| Step | What you do |
|---|---|
| 1 | You update Hermes to v0.22.3 or newer with hermes update, or through the Store for the Windows Store app. |
| 2 | You open Settings, then Providers, then Local Models, and click Install runtime. |
| 3 | You pick Nemotron 3.5 Bolt from the catalog and click Download. |
| 4 | You read the memory fit colour, then click Use. |
In a terminal, the docs say hermes model then Local models runs the same setup.
The catalog file in the Hermes repo at the v0.22.4 tag shows Hermes pulls an IQ2 model file of about 22 GB, a vision file of about 1.7 GB and a LiLiCorr draft model of about 1 GB.
The same entry is set so Hermes doesn't auto-recommend it, so you choose it yourself.
A tip from my Solar Mini 4 video below is to keep one Hermes profile per model, so you can switch brains and compare them without breaking your main setup.
If you'd like help working out where agents fit your business and your search traffic, 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.
Every feature of Bolt in Hermes
| Feature | How it works |
|---|---|
| One-click download | The Local Models catalog downloads the model, vision file and draft model together. |
| Memory fit check | The docs say each row shows green when it fits your GPU, amber when it spills into RAM and red when it's too big. |
| Growing context | The docs say local models start at a window that fits your GPU and grow toward their maximum as the chat needs it. |
| Idle unloading | The docs say idle models unload after 15 minutes and reload on the next message. |
| Pause and resume | The docs say downloads pause and resume, and Ctrl+C pauses a terminal download. |
| Own llama-server | The docs say Hermes detects a llama-server you already run, with port 8080 as the default probe. |
| Vision | The Hermes catalog describes Bolt as a model that sees images. |
| Speed trick | The DFlash draft model guesses several tokens ahead and the main model checks them. |
Super VL through NVIDIA NIM
If your hardware can't run Bolt, its parent model is the cloud route.
NVIDIA's card lists Nemotron 3.5 Super VL 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 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, the docs say to set NVIDIA_BASE_URL=http://localhost:8000/v1 instead.
Limits to know first
NVIDIA's recipes start at 32GB RTX cards and 64GB Spark machines, so most laptops are out.
The Hermes v0.22.3 notes say the Bolt support is undocumented on purpose until v0.23.0.
The Hermes Local Models docs still say Hermes never offers builds below 4-bit, while Bolt is NVIDIA's own 2-bit build.
NVIDIA's blog lists the checkpoints under the name "Nemotron 3.5 Super VL 74B-A9B", while its model cards call them Bolt.
I haven't run Bolt myself, so everything here comes from NVIDIA's pages, the Hermes docs, the Hermes catalog file and the release notes.
What NVIDIA's own benchmarks say
These numbers come from NVIDIA's model card and blog, not from GoldieBench.
| Benchmark | Nemotron 3.5 Bolt (IQ2) | Qwen3.6-35B-A3B | Gemma 4 26B-A4B |
|---|---|---|---|
| Terminal-Bench 2.1 | 50.98 | 44.38 | 37.22 |
| BrowseComp | 66.75 | 48.74 | 26.30 |
| GPQA Diamond | 86.43 | 83.40 | 79.61 |
| PinchBench | 87.82 | 88.07 | 74.70 |
| Coding speed test (tokens per second) | 83.99 | 96.63 | 100.56 |
NVIDIA ran the speed test with llama.cpp on an RTX Spark laptop with 128GB.
On NVIDIA's own figures, Bolt leads on agent tasks while Qwen and Gemma lead on raw speed.
Is Nemotron 3.5 Bolt on GoldieBench?
No, it isn't, and no Nemotron model is on the board at all yet.
GoldieBench gives every model the same one-shot prompt to build a game, a page, a simulation or a visual, then scores the rendered result from 0 to 10.
That's a different test from NVIDIA's agent benchmarks, so a strong Terminal-Bench score doesn't tell you how Bolt would build a game in one shot.
NVIDIA's two comparison models, Qwen3.6-35B-A3B and Gemma 4 26B-A4B, aren't on the board either.
Here are the local models that are on the live board, so you can see where the local bar sits today.
| Model | On GoldieBench | Live score | Note |
|---|---|---|---|
| Qwable 5 27B Coder | Qwable 5 27B Coder | 7.14 average | It is a fine-tune of Qwen3.6-27B, with 41 scored tasks. |
| Agents-A1 | Agents-A1 | 4.83 average | It is an open-weights model that runs locally, with 45 scored tasks. |
| Gemma 4 12B · MLX | Gemma 4 12B · MLX | 3.98 average | It is Google's small local model, with 42 scored tasks. |
| Laguna XS 2.1 | Laguna XS 2.1 | 3.93 average | It is Poolside's small model, with 42 scored tasks. |
| Qwythos 9B | Qwythos 9B | 2.98 average | It is a 9B local model, with 42 scored tasks. |
On those numbers, the 27B-class local model scores far above the small ones, which fits NVIDIA's argument that bigger models squeezed into the same memory beat small ones.
Bolt is a 74B model, so it's the obvious next local candidate for the board, but I won't guess a score before it's tested.
If you're choosing a local brain today, my post on the best local model for Hermes Agent uses the same live board.
You can also check every local entry on the local models board.
Who should use Nemotron 3.5 Bolt in Hermes
Operators with an RTX card that has 32GB or more, or an RTX or DGX Spark machine, should try it first.
Businesses that need a private agent for documents, screenshots and research should look at it closely.
Anyone on a smaller machine should pick a smaller local model or use Super VL through NVIDIA NIM.
For the release that made local models one click, my Hermes Agent v0.22 post covers the rest of the update.
If you want my team to show you how to turn AI agents and content into Google and AI search traffic, book a free SEO strategy session.
Also On Our Network
- Agent Operatorsthe four-step Nemotron 3.5 Bolt Hermes setup
- AI Income Desksix ways a business can use Nemotron 3.5 Bolt in Hermes
- AI Tool Verdictevery Nemotron 3.5 Bolt spec, checked and scored
- agentos.guidethe Nemotron 3.5 Bolt quick start for an Agent OS
- agentos.guidemy best local model for Hermes Agent guide on agentos.guide
If you want a clear plan for getting your business found in Google and AI search, book your free strategy session with my team.
Check your card, run it from Local Models and watch the board for a real score, and you'll know exactly where your Nemotron 3.5 Bolt Hermes setup stands.
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 card says it has 74B total and 9B active parameters, uses native 2-bit quantisation to fit in 24GB, 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.
Is Nemotron 3.5 Bolt on GoldieBench?
No. No Nemotron model is on the GoldieBench leaderboard yet. The local models with live scores include Qwable 5 27B Coder at 7.14, Agents-A1 at 4.83 and Gemma 4 12B MLX at 3.98.
Does Hermes ship Nemotron 3.5 Bolt as its own model?
No. Bolt is NVIDIA's model, not a Hermes model. Hermes v0.22.3 added it to the Local Models catalog, which downloads it from NVIDIA's own Hugging Face repo.
When was Nemotron 3.5 Bolt released?
NVIDIA's model cards and technical blog give 15 October 2026, and Hermes added it to the local runtime in v0.22.3 on 21 October 2026.
Keep reading
Hermes Herald AI Agent: Which Brain To Run It On, With Real GoldieBench Scores
GEO Examples: Four Real AI Overview Wins, The Three Moves Behind Them, And The Models Writing The Answers
Best Generative Engine Optimization Agency: Seven Picks By Region, And How The Models Behind AI Answers Score🌐 Sister-site take: read this on agentos.guide