Agents / AtomicChat · NVIDIA base model

NVIDIA Nemotron 3.5 Bolt 74B-A9B Uncensored

AtomicChat’s uncensored GGUF builds of NVIDIA’s Nemotron 3.5 Bolt 74B-A9B agent model, made from the full-precision release with the refusal direction removed. Files from 21.9 GB to 51.7 GB for 24–64 GB machines.

Verified sourceSource checked 11/17/2026
LOCALOPEN WEIGHTS
Freshness 87 out of 100. Repository age: 37 days old, Oct 15, 2026. Measures age, not quality.
Popularity 57 out of 100. 29,863 downloads in 30 days; 23 lifetime likes. Checked Nov 21, 2026. Measures adoption, not quality.
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.

Check the date source
Popularity · adoption

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.

round(80 × min(1, log10(1 + downloads) / 7) + 20 × min(1, log10(1 + likes) / 5))

Check the statistics source
Scores as of Nov 21, 2026. Statistics checked Nov 21, 2026. Statistics are marked for refresh after 72 hours.Full scoring method
WORKFLOWS FOR THIS MODEL 1

What do you want to do?

Choose a task. Get the software, files, settings, and steps to run it.

MAKE IT PERSONALChoose your hardware

Open a guide for its software, hardware requirements, and setup steps.

At a glance

Parameters
74B total · about 9B active per token
Architecture
Hybrid Mamba-Transformer mixture of experts (nemotron_h_moe)
Context length
AtomicChat’s example uses 32,768 and reports a passing retrieval test at 30K tokens · the agent guide starts at 65,536
License
OpenMDW-1.1, as the original
Software
llama.cpp
MY HARDWARE

Will it run on your machine?

Save your machine to see a personalized rating and its reasoning.

Add my hardware

Best for

  • Run a local Hermes agent that reads and writes files with your approval.
  • Private chat and reasoning with fewer refusals on a 64 GB workstation or Mac.
  • Coding and tool-use tasks that need a larger model than the 24–32 GB picks.

Tradeoffs

  • AtomicChat’s numbers (JailbreakBench refusals 97.5% to 0.0%, MMLU 81.35% to 81.75%, 20 of 20 valid tool calls) are publisher results, not YouRunAI tests.
  • With thinking on, the AD-IQ2_XS and AD-IQ1_M_S files each still refused 2 of 30 harmful prompts in AtomicChat’s check; AD-Q4_K_M was not checked.
  • Image input was not tested on these files.
  • The smaller the file, the further it drifts from the original: 93% same top answer for AD-Q4_K_M, 86% for AD-IQ2_XS, 79% for AD-IQ1_M_S.
  • Not an NVIDIA release. The card states deployments need their own access controls.

Ways to run it

llama.cpp · Windows, macOS, Linux

AtomicChat’s example file and settings. The Hermes agent guide on this page raises the context to 65,536.

llama-server --hf-repo AtomicChat/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-Uncensored-GGUF --hf-file NVIDIA-Nemotron-3.5-Bolt-74B-A9B-Uncensored-AD-IQ2_XS.gguf -ngl 99 -c 32768 --jinja
Official instructions

llama.cpp · AD-Q4_K_M · Windows, macOS, Linux

Two-part file; llama.cpp loads the second part on its own.

llama-server -hf AtomicChat/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-Uncensored-GGUF:Q4_K_M -ngl 99 -c 32768 --jinja
Official instructions

What to expect

The pick for 64 GB of memory. Start with AD-Q4_K_M in llama.cpp on a 64 GB machine with other apps closed, and fall back to AD-IQ2_XS on 48 GB or when memory runs short.

Explore its uses

UNCENSOREDAGENTSTOOL CALLINGREASONINGMIXTURE OF EXPERTSGGUFNEMOTRON
REFERENCE LIBRARY

Sources and files

Primary sources and files behind this model page.

3 SOURCES

Keep exploring