The Nemotron 3.5 Bolt Hermes setup puts NVIDIA's new 74B-parameter local model in charge of Hermes Agent, and my verdict is that it's the most interesting local brain Hermes has added, as long as you own a serious NVIDIA machine.
NVIDIA's model card describes it as a vision-language model with 74B total and 9B active parameters, squeezed into a native 2-bit build to fit in 24GB of memory.
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.
The catches are a high hardware bar, Hermes docs that haven't caught up yet, and benchmark numbers that so far come only from NVIDIA.
Below I score every part of it from NVIDIA's own pages, the Hermes docs, the Hermes catalog file and the official release notes, then give you my final verdict.
If you want my team to look at how AI agents and AI search can grow your business, book a free SEO strategy session with my team.
The Nemotron 3.5 Bolt Hermes Verdict In One Table
These are my scores out of 5, based on NVIDIA's model cards and blog plus the Hermes docs, catalog and release notes rather than my own run.
| Feature | Score | Why it got that score |
|---|---|---|
| Size for its brains | 5/5 | It fits a 74B-parameter model with 9B active into a 24GB memory target. |
| Vision input | 4.5/5 | It reads images and MP4 videos up to two minutes, which suits agents that look at screens. |
| Context window | 4.5/5 | It supports up to 256K tokens, and NVIDIA's bigger recipe runs 262,144 tokens. |
| Agent benchmarks | 4/5 | NVIDIA's table shows strong agent scores, but they're the maker's own numbers. |
| Speed | 3.5/5 | It hit 83.99 tokens per second on NVIDIA's test, behind Qwen3.6 35B-A3B and Gemma 4 26B-A4B. |
| Licence | 4.5/5 | OpenMDW-1.1 allows commercial or non-commercial use, according to the card. |
| Hermes integration | 4.5/5 | It's one click in Local Models, with the vision file and draft model set up for you. |
| Hardware reach | 2.5/5 | NVIDIA's recipes need 32GB-plus RTX cards or 64GB-plus Spark machines. |
| Documentation | 2.5/5 | Nous says the Bolt support is undocumented until v0.23.0, and the docs still say no builds below 4-bit. |
| Cloud fallback | 4.5/5 | Its parent, Super VL, works in Hermes through the NVIDIA NIM provider. |
| Overall | 4/5 | It's a big brain in a small footprint, held back by the hardware it needs and thin independent testing. |
🔥 Want my exact Is Nemotron 3.5 Bolt In Hermes Any Good setup, step by step? Inside the AI Profit Boardroom you get step-by-step video trainings, my installable Agent OS and four coaching calls a week with 3,400+ members building real automations. → Join the AI Profit Boardroom
What Nemotron 3.5 Bolt Is
NVIDIA's full name for it is NVIDIA Nemotron 3.5 Bolt 74B-A9B 24GB.
The 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.
| Spec | What NVIDIA's model card says |
|---|---|
| Parameters | It has 74B total and 9B active parameters. |
| Architecture | It is a Mamba-2 and Transformer hybrid LatentMoE with a C-RADIO v4-H vision encoder. |
| Precision | It uses IQ2, which is 2-bit quantisation with quantisation-aware training. |
| Context | It supports up to 256K tokens. |
| Inputs and outputs | It takes text, images and video, and it outputs text. |
| Speculative decoding | It uses DFlash with a separate LiLiCorr drafter. |
| Reasoning | Thinking can be switched on or off through the chat template. |
| Recommended sampling | NVIDIA recommends a temperature of 1.0 and a top_p of 0.95. |
| Licence | It is governed by the OpenMDW License Agreement, version 1.1. |
| Release date | The cards give 15 October 2026. |
NVIDIA's blog says the design goal was to fit into 24GB and decode at a speed of at least 80 tokens per second.
How To Set Up 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.
- Update Hermes to v0.22.3 or newer with
hermes update, or through the Store if you use the Windows Store app. - Open Settings, then Providers, then Local Models, and click Install runtime.
- Pick Nemotron 3.5 Bolt from the catalog and click Download.
- Check the memory fit, then click Use.
In a terminal, the docs say hermes model then Local models does the same job.
🔥 Want my exact Hermes and local model setup? Inside the AI Profit Boardroom, I've got a full Hermes section with video tutorials and the installable Agent OS I run every day. You'll be learning with 3,400+ members and four coaching calls a week. → Get access here
🚀 Ready to put Is Nemotron 3.5 Bolt In Hermes Any Good to work in your business? Inside the AI Profit Boardroom you get step-by-step video trainings, my installable Agent OS and four coaching calls a week with 3,400+ members building real automations. → Join the AI Profit Boardroom
Feature By Feature: Nemotron 3.5 Bolt In Hermes Reviewed
Size for its brains scores 5 out of 5
This is the headline feature, and it earns full marks.
NVIDIA quantised all experts to IQ2_XXS at 2.0625 bits per element, kept the Mamba layers in NVFP4, and kept the embedding and output layers at Q8_0.
NVIDIA's blog says its tests show that compressing a large model to fit a 24GB budget beats training a smaller one from scratch, and that larger networks at 1 to 2 bits are more Pareto-optimal than smaller ones at higher bits.
That's a big deal for local agents, because memory is usually what caps how clever a home model can be.
Vision input scores 4.5 out of 5
The card says Bolt takes RGB images in JPEG or PNG, plus MP4 video up to two minutes.
The Hermes catalog describes it as a model that sees images, and it downloads the vision projector file automatically.
For an agent that reads screenshots and web pages, that's exactly the right skill.
Context window scores 4.5 out of 5
The card says the model supports up to 256K tokens, shared between the prompt, images, history, tool results and the reply.
NVIDIA's bigger llama.cpp recipe runs a 262,144-token window, and the smaller one runs 65,536 tokens.
The Hermes docs say local models start at a context window that fully fits your GPU and grow as the conversation needs more room.
Agent benchmarks score 4 out of 5
These are NVIDIA's own published numbers.
| Benchmark | Nemotron 3.5 Bolt (IQ2) | Qwen3.6-35B-A3B | Gemma 4 26B-A4B | Nemotron 3.5 Lightning 30B-A3B |
|---|---|---|---|---|
| Terminal-Bench 2.1 | 50.98 | 44.38 | 37.22 | 23.46 |
| BrowseComp | 66.75 | 48.74 | 26.30 | 36.81 |
| MCP-Atlas | 52.20 | 51.20 | Not reported | 46.4 |
| GPQA Diamond | 86.43 | 83.40 | 79.61 | 75.57 |
| PinchBench | 87.82 | 88.07 | 74.70 | 83.43 |
| AA-Omniscience accuracy | 16.90 | 19.47 | 22.17 | 16.63 |
NVIDIA notes that Qwen's own reported MCP-Atlas score is 62.8, while NVIDIA measured 51.2 on the 500 public tasks.
NVIDIA says Bolt is first on PinchBench, second on WebArena and second on BrowseComp among models with a similar memory footprint.
I've knocked a point off because nobody outside NVIDIA has published a matching set of results that I could verify, and Qwen and Gemma still win a couple of rows.
Speed scores 3.5 out of 5
NVIDIA's blog measured speed with llama.cpp on an RTX Spark laptop with 128GB.
| Model | Coding speed test (tokens per second) |
|---|---|
| Nemotron 3.5 Bolt with LiLiCorr | 83.99 |
| Qwen3.6 35B-A3B with MTP | 96.63 |
| Gemma 4 26B-A4B with MTP | 100.56 |
| Qwen3.8 27B with DFlash | 28.26 |
Bolt hits NVIDIA's 80 tokens per second target, but it isn't the fastest model in its own table.
Its pitch is more intelligence per GB, not raw speed.
Licence scores 4.5 out of 5
The card says the model is ready for commercial or non-commercial use.
Use is governed by the OpenMDW License Agreement, version 1.1, so read it before you build a product on top.
Hermes integration scores 4.5 out of 5
The Hermes catalog file at the v0.22.4 tag lists Nemotron 3.5 Bolt with NVIDIA's own GGUF repo as the source.
It pulls the IQ2 model file of about 22 GB, a vision file of about 1.7 GB and a LiLiCorr draft model of about 1 GB for DFlash.
The entry is set so Hermes doesn't auto-recommend it, which means you choose it yourself.
That's sensible, because Hermes only recommends models that fit fully in GPU memory at a good speed.
Hardware reach scores 2.5 out of 5
NVIDIA's GGUF card gives two recipes, one for RTX Spark 128GB or RTX cards with 48GB+, and one for RTX Spark 64GB or RTX cards with 32GB+.
It lists Ada Lovelace and Blackwell GPUs on Windows and Linux, and it doesn't mention Macs.
The "24GB" in the name is the model's memory target, not the size of card the recipes ask for.
If you'd like my team to look at where AI agents and content fit your growth plan, grab a free strategy session with my team and we'll go through your site and the keywords you're closest to winning.
Documentation scores 2.5 out of 5
The Hermes v0.22.3 notes say this part of the window is undocumented on purpose, and the full notes ship with v0.23.0.
The Local Models docs still say Hermes never offers builds below 4-bit, which doesn't match a 2-bit model in the catalog.
NVIDIA's blog also lists the checkpoints as "Nemotron 3.5 Super VL 74B-A9B", while its model cards call them Bolt, so the naming is messy.
Cloud fallback scores 4.5 out of 5
NVIDIA's card for Nemotron 3.5 Super VL lists 120B total and 12B active parameters and up to 1M tokens of context.
The Hermes docs run it through the NVIDIA NIM provider with this command, once NVIDIA_API_KEY is in ~/.hermes/.env.
hermes chat --provider nvidia --model nvidia/nemotron-3.5-super-vl-120b-a12b
The Hermes v0.22.1 and v0.22.2 notes added Super 3.5 default parameters and put Nemotron on the NIM default model list.
Nemotron 3.5 Bolt vs The Other Hermes Local Brains
The Hermes v0.22.0 notes name Qwen3.8 27B, Qwen3.8 Flash Next, Qwen3.6 35B-A3B and DeepSeek V4 Flash in the local catalog.
| Model | Where it fits, based on official sources |
|---|---|
| Nemotron 3.5 Bolt | It is NVIDIA's 74B 2-bit build for 32GB-plus RTX cards and Spark machines. |
| Qwen3.6 35B-A3B | It beat Bolt on PinchBench and on speed in NVIDIA's own tables. |
| Qwen3.8 27B | It was much slower than Bolt on NVIDIA's speed test. |
| DeepSeek V4 Flash | The Hermes catalog describes it as a frontier-class model for machines with 128GB+ memory. |
The Hermes docs say a GPU with 8 GB or more runs the small catalog models comfortably, and 16 GB or more runs the 27B to 35B models at high quality.
Limits And What It Doesn't Do
It doesn't run on ordinary laptops, because NVIDIA's recipes need big NVIDIA memory.
It doesn't have independent benchmark coverage yet, so every score in this review's benchmark table is NVIDIA's.
It isn't on GoldieBench, because no Nemotron model is on my leaderboard yet.
I haven't run Bolt myself, so this review 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 before, which I showed in my Gemma 4 video below.
🔥 Want to build real automations on top of a local Hermes brain? The AI Profit Boardroom has the Agent OS, the Hermes training and four weekly coaching calls where you can show me your setup. → Join the AI Profit Boardroom
Final Verdict
I give Nemotron 3.5 Bolt in Hermes 4 out of 5.
It packs a 74B-parameter model with vision into a 24GB target under an open licence, and Hermes makes it one click.
It loses a point because NVIDIA's own recipes need serious hardware, the Hermes docs haven't caught up, and the benchmark story is still NVIDIA's alone.
If you're choosing a brain more broadly, my best Hermes Agent model ranking covers the cloud options, and my Solar Mini 4 Hermes review covers another free model.
Related Reading
- My best Hermes Agent model ranking compares the brains you can plug into Hermes.
- My Hermes Agent v0.22 review scores the release that brought one-click local models.
- My Hermes Agent Microsoft Store review scores the easiest way to run Hermes on a Windows PC.
Also On Our Network
- 🌐 The four-step Nemotron 3.5 Bolt Hermes setup walkthrough
- 🌐 Six ways a business can use Nemotron 3.5 Bolt in Hermes
- 🌐 The Nemotron 3.5 Bolt quick start for an Agent OS
- 🌐 Whether any Nemotron model is on GoldieBench yet
FAQ: Nemotron 3.5 Bolt Hermes Verdict
Is Nemotron 3.5 Bolt good for Hermes Agent?
On NVIDIA's own numbers it is strong for agent work, scoring 50.98 on Terminal-Bench 2.1 and 66.75 on BrowseComp.
It needs a powerful NVIDIA machine, and independent tests are still thin.
What are the Nemotron 3.5 Bolt specs?
NVIDIA lists 74B total and 9B active parameters, IQ2 2-bit quantisation, up to 256K tokens of context and text, image and video inputs.
It is licensed under OpenMDW-1.1.
Does it run on a 24GB graphics card?
NVIDIA designed the model to fit in 24GB, but its llama.cpp recipes are written for RTX Spark with 64GB or 128GB and RTX cards with 32GB or 48GB and up.
Hermes shows a memory fit check for your own machine before you download.
Is it faster than Qwen3.6 35B-A3B?
Not on NVIDIA's speed test, which shows 83.99 tokens per second for Bolt against 96.63 for Qwen3.6 35B-A3B.
Is Nemotron 3.5 Bolt on GoldieBench?
No, because no Nemotron model is on the GoldieBench leaderboard yet.
If you'd like my team to help your business win Google and AI search traffic with agents and content, book your free SEO strategy session here.
📺 Video notes + links to the tools 👉
🎥 Learn how I make these videos 👉
🆓 Get a FREE AI Course + Community + 1,000 AI Agents 👉
About Julian
I'm Julian Goldie, an SEO entrepreneur, author and founder of the AI Profit Boardroom, which has 3,400+ members.
I help business owners scale with AI agents, automation and SEO.
- I've built a 7-figure SEO and link-building agency, Goldie Agency, with 70+ people in Bangkok.
- I've grown a YouTube channel to 400,000+ subscribers.
- I wrote the book "Link Building Mastery", which covers nine link building strategies.
- I run GoldieBench, where I score AI models on real one-shot builds.
→ Get my best AI training inside the AI Profit Boardroom
If you own a 32GB-plus RTX card or a Spark machine, my verdict is to try the Nemotron 3.5 Bolt Hermes setup this week.
