Workflows/Build a local AI agent with Nemotron 3.5 Bolt 74B-A9B Uncensored
Agents / WORKFLOW
Build a local AI agent with Nemotron 3.5 Bolt 74B-A9B Uncensored
Run AtomicChat’s Nemotron 3.5 Bolt 74B-A9B Uncensored with Ollama (29 GB build) or llama.cpp (52 GB two-part build) on Windows, Mac or Linux, connect Hermes Agent, and have it read two files and write a checked review.
Verified sourceSource checked 11/17/2026
LOCAL
Freshness 87 out of 100. Repository age: 37 days old, Oct 15, 2026. Measures age, not quality.87/ 100Freshness37 days oldRepository age · Oct 15, 2026
Popularity 57 out of 100. 29,863 downloads in 30 days; 23 lifetime likes. Checked Nov 21, 2026. Measures adoption, not quality.57/ 100Popularity29.9K downloads / 30d23 lifetime likes · Checked Nov 21, 2026
What do these scores mean?+
Two ways to compare models. Neither score measures answer quality, benchmark performance, or whether a model fits your machine.
Freshness · release age
A new release starts at 100. The score halves every 180 days: 100 × 2^(−age in days / 180). When the release date is unavailable, a repository date is a fallback—not proof of when the model was released. Missing dates stay unscored.
Downloads over the past 30 days carry 80% of the score; lifetime likes carry 20%. Both use a logarithmic scale, capped at 10 million downloads and 100,000 likes. If likes are unavailable, downloads carry all the weight. Missing download statistics stay unscored.
Scores as of Nov 21, 2026. Statistics checked Nov 21, 2026. Statistics are marked for refresh after 72 hours.Full scoring method →
Before you begin
Hardware
Estimated from the published file sizes plus runtime and context headroom at 65,536 tokens. Not measured by YouRunAI. Bolt is a mixture-of-experts model that uses about 9B of its 74B parameters per token, so keeping part of it in system RAM costs less speed than it would with a dense model.
Machine
AD-Q4_K_M · 51.7 GB
AD-IQ2_XS · 29.1 GB
Apple Silicon · 64 GB unified memory
Recommended start; the download alone is 51.7 GB, so close everything else
Comfortable fallback
Apple Silicon · 96 GB unified memory or more
Comfortable
Comfortable
Apple Silicon · 48 GB unified memory
Cannot fit
Recommended for 48 GB; close other apps
NVIDIA · 80 GB VRAM or more
Recommended
Comfortable
NVIDIA · 48 GB VRAM + 64 GB RAM
Does not fit on the GPU alone; lower -ngl so the rest sits in RAM
Recommended
NVIDIA RTX 5090 · 32 GB VRAM + 64 GB RAM
Partial offload
Nearly fits; lower -ngl a little
NVIDIA · 24 GB VRAM + 64 GB RAM
Mostly in RAM; slow
Partial offload: lower -ngl so the rest sits in RAM
CPU only · 64 GB RAM
Too tight
Loads; slow, short tasks only
AtomicChat also publishes a 21.9 GB AD-IQ1_M_S file for 24 GB machines; it is the furthest from the original in AtomicChat’s tests (79% same top answer), so this guide leaves it out. NVIDIA DGX Spark is not covered by the Linux tab here, which uses the x86 CUDA Docker image.
LOCAL AGENT · HERMES · WINDOWS · MAC · LINUXSOURCE-LINKED SETUP
Build a local AI agent with Nemotron 3.5 Bolt 74B-A9B Uncensored
Run AtomicChat’s GGUF in Ollama or llama.cpp, point Hermes at it, and have the agent read two project files and write review.md. You approve each file action. Bolt was built by NVIDIA for local agents, and AtomicChat reports tool calls unaffected by its edit.
Choose a package Run its commands Check the result
01
BEFORE YOU BEGIN
Choose your package
Select the weights you want. The instructions, downloads and settings change with your choice.
The publisher has not provided a complete memory target for this task. Check its hardware guidance before starting.
The 51.7 GB download is only the weights, so a 64 GB machine has little to spare once the runtime and a 65,536-token context are added: close other apps, and switch to AD-IQ2_XS if loading runs out of memory. Not a YouRunAI measurement.
Freshness 87 out of 100. Repository age: 37 days old, Oct 15, 2026. Measures age, not quality.87/ 100Freshness37 days oldRepository age · Oct 15, 2026
Popularity 57 out of 100. 29,863 downloads in 30 days; 23 lifetime likes. Checked Nov 21, 2026. Measures adoption, not quality.57/ 100Popularity29.9K downloads / 30d23 lifetime likes · Checked Nov 21, 2026
What do these scores mean?+
Two ways to compare models. Neither score measures answer quality, benchmark performance, or whether a model fits your machine.
Freshness · release age
A new release starts at 100. The score halves every 180 days: 100 × 2^(−age in days / 180). When the release date is unavailable, a repository date is a fallback—not proof of when the model was released. Missing dates stay unscored.
Downloads over the past 30 days carry 80% of the score; lifetime likes carry 20%. Both use a logarithmic scale, capped at 10 million downloads and 100,000 likes. If likes are unavailable, downloads carry all the weight. Missing download statistics stay unscored.
Set up Nemotron 3.5 Bolt Uncensored · AD-Q4_K_M on your machine
Choose your computer first, then its GPU path. The instructions below update to match your selection.
CHOOSE YOUR SETUPWindows
01
Install llama.cpp with winget
Open PowerShell and install the official llama.cpp release with winget, then close and reopen PowerShell so llama-server is on your PATH. AtomicChat’s card says upstream llama.cpp with nemotron_h_moe support runs these files without a patch, so use a current release. The second command prints your build number, and the third lists the GPUs llama.cpp can use. If only the CPU appears on an NVIDIA or AMD PC, download the CUDA or HIP Windows build from the official releases page instead and run the same commands from its folder.
Download Nemotron 3.5 Bolt Uncensored · AD-Q4_K_M and start the local server
In PowerShell, run this command and leave the window open. The first run downloads AtomicChat’s 51.7 GB two-part AD-Q4_K_M download; later runs reuse it. The server listens only on this computer at http://127.0.0.1:8080, names the model nemotron-3.5-bolt-uncensored, and starts at 65,536 tokens of context with one request slot (-np 1), so the whole context goes to the agent. Hermes documents a 64K minimum context for local models, which is why this guide does not start lower. The --no-mmproj flag skips the vision file: AtomicChat did not test image input on these files, and this guide only works with text. The chat template is inside the file, and --jinja applies it. AtomicChat’s card gives no sampling settings, so the server defaults are used. Wait until the log says the server is listening.
Open a second PowerShell window on your own computer and run the official Hermes installer. Close and reopen PowerShell afterwards so the hermes command is available. This native installer does not need WSL.
Run these custom-provider settings in that PowerShell window. They are the same three settings Hugging Face shows in the Hermes Agent section of GGUF model pages, with this server’s address and the model name nemotron-3.5-bolt-uncensored. Hermes then sends every request to llama.cpp on this computer.
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default nemotron-3.5-bolt-uncensored
05
Create two small project files
In the same PowerShell window, create a disposable folder with a launch note and a checklist. The checklist deliberately has no database backup before the migration; the agent should find that gap from the files.
PowerShell
New-Item -ItemType Directory -Force "$HOME\agent-review" | Out-Null
Set-Location "$HOME\agent-review"
@"
Launch day: Thursday 14 May
Owner: Priya
Database migration runs at 09:00 before the app deploy.
"@ | Set-Content notes.md
@"
1. Freeze merges
2. Run the database migration
3. Deploy the app
4. Announce the release
"@ | Set-Content checklist.md
06
Give Hermes a file-based task
From inside the agent-review folder, start Hermes and paste the task below. Hermes asks before running commands or writing files; read each request before approving it. Approve the file reads and the single write of review.md. A model with reduced refusals will not push back on a risky action, so the approval prompt is your safeguard: keep it on.
Setting
Value
Task to paste into Hermes
Read notes.md and checklist.md in this folder. Write review.md containing: the launch day and owner, a corrected numbered checklist that adds a verified database backup before the migration, and one sentence explaining the risk of migrating without a backup. Cite the file each fact came from. Do not change notes.md or checklist.md.
Open review.md. It should name Thursday 14 May and Priya from notes.md, add a backup step before “Run the database migration”, and explain the risk. Check that Hermes’ activity shows real file reads and one write; an answer only in the chat window is not a completed agent run. Type /exit to leave Hermes and press Ctrl+C in the server window when you are done. To use the agent again, run the same server command (nothing is downloaded twice) and start hermes in any project folder.
review.md names the launch day and owner from notes.md, adds a verified backup before the migration, and Hermes’ activity shows the file reads and one approved write.
AtomicChat reports 20 of 20 valid tool calls after its edit and shows these Hermes custom-provider settings on the model’s Hugging Face page. Those are publisher results; YouRunAI has not run this exact task with these files, so inspect the output rather than trusting the chat summary. Sources checked November 17, 2026.
Check that the server was started with --jinja, the model name is nemotron-3.5-bolt-uncensored, and you started Hermes inside the agent-review folder. A chat-only answer is not a passing run.
ollama pull fails, or the model says its architecture is unknown.
Update Ollama to 0.40.16 or newer (re-run the installer, or Check for Updates from its icon), restart it, and run the command again. ollama pull resumes an interrupted download.
Ollama runs out of memory, or ollama ps shows the model partly on the CPU and it is very slow.
Choose a smaller package if there is one, or move Ollama’s context slider down to 32k. If it runs out of memory as soon as Hermes starts, run hermes config set model.context_length 65536 so Hermes does not ask for the model’s full maximum context.
The server stops with an out-of-memory error, or loading never finishes.
Close other apps that use the GPU or memory, or switch to the AD-IQ2_XS package. On a PC whose GPU is smaller than the file, replace -ngl 99 with a lower number such as -ngl 20 so part of the model stays in system RAM; it will be slower. Lowering -c below 65536 saves memory, but Hermes may then warn that the context is too small for agent work.
llama-server says it does not recognise the model architecture.
These files use the nemotron_h_moe architecture, which older llama.cpp builds cannot load. Update llama.cpp (winget upgrade llama.cpp, brew upgrade llama.cpp, or docker pull the server image again), check llama-server --version, and try again.
The installer finishes but hermes is not found.
Close and reopen the terminal. Follow any PATH or dependency message the installer printed.
REFERENCE LIBRARY
Sources and files
Original instructions, model files, and compatibility notes behind this setup.