Llama-3.1-Nemotron-Ultra-253B
A 253B reasoning model derived from Llama-3.1-405B-Instruct
NVIDIA's Llama-3.1-Nemotron-Ultra-253B is a dense decoder model derived from Meta's Llama-3.1-405B-Instruct through neural architecture search. Trained between November 2024 and April 2025, it targets reasoning, chat, retrieval and tool use, and runs on one 8xH100 node.
- 253B parameters, derived from Llama-3.1-405B-Instruct
- Fits on a single 8xH100 node for inference
- Neural architecture search yields irregular, non-repeated blocks
- 128K-token context window
- Trained November 2024 to April 2025
What it is
Llama-3.1-Nemotron-Ultra-253B-v1 is a large language model from NVIDIA, built as a derivative of Meta's Llama-3.1-405B-Instruct. It carries 253 billion parameters and is aimed at reasoning work as well as ordinary chat. NVIDIA positions it for agent systems, chatbots, retrieval-augmented generation and tool calling, and says it is cleared for commercial use. A context window of 128K tokens is supported, and the whole model can be served on a single eight-GPU H100 node. The company presents it as a balance point between answer quality and the cost of running the model.
How it came about
The model belongs to the Llama Nemotron Collection, alongside Llama-3.1-Nemotron-Nano-8B-v1 and Llama-3.3-Nemotron-Super-49B-v1. NVIDIA reports that training ran from November 2024 to April 2025. Pretraining data inherited from the reference model stops at 2023. A technical report and a developer blog describe the work, and NVIDIA invites users to vote on future features for the Nemotron line. No release date is given.
Inside the model
The starting point is Llama-3.1-405B-Instruct, reshaped with a neural architecture search method that produces irregular blocks rather than a repeated pattern. Some blocks drop attention altogether or swap it for one linear layer, feed-forward layers vary in their expansion ratio, and runs of skipped attention let several feed-forward layers be merged into fewer, wider ones. Variants of each block are generated and then searched to hit a throughput and memory target with the smallest quality loss. Knowledge distillation over 65 billion tokens restores capability, followed by 88 billion tokens of further pretraining. Post-training covers maths, code, reasoning, chat and tools through supervised tuning plus several GRPO reinforcement learning rounds. The company says Qwen also improved the model.
What it can do
Intended for English and coding languages, with German, French, Italian, Portuguese, Hindi, Spanish and Thai also supported. Reasoning can be switched on or off through the system prompt; NVIDIA suggests temperature 0.6 and top-p 0.95 when it is on, and greedy decoding when it is off. When reasoning is on and none is needed, the model emits a marker that NVIDIA says is expected. No benchmark figures are published here.
How to use it
Weights are downloadable under the NVIDIA Open Model License, with the Llama 3.1 Community License also applying, and NVIDIA states the model is ready for commercial use. It can be run through Hugging Face Transformers or served with vLLM, and a preview API is offered on NVIDIA's build site. Linux is the preferred operating system.
Why it matters
NVIDIA released its Llama-Nemotron post-training dataset publicly, saying the aim is openness in how such models are built and improved. The architecture search and vertical compression techniques behind the model are described in published papers, including work on distillation-based search and on fusing feed-forward layers.
Questions people ask about Llama-3.1-Nemotron-Ultra-253B
Where does this model come from?
It is a derivative of Meta's Llama-3.1-405B-Instruct, produced by NVIDIA using neural architecture search to cut the parameter count to 253 billion while keeping quality close to the reference model.
What hardware do I need to run it?
NVIDIA says the model fits on one node with eight H100 GPUs for inference. The reduced memory footprint is meant to lower the number of GPUs required in a data centre compared with the 405B reference model.
Can I use it commercially?
Yes. NVIDIA states the model is ready for commercial use. Weights fall under the NVIDIA Open Model License, with the Llama 3.1 Community License Agreement also applying to the underlying Llama 3.1 components.
How do I turn reasoning on and off?
A system prompt controls it. NVIDIA recommends temperature 0.6 and top-p 0.95 with reasoning enabled, greedy decoding with it disabled, and putting all instructions in the user message rather than adding extra system prompts.
What languages does it handle?
It is aimed at English and coding languages, and also supports German, French, Italian, Portuguese, Hindi, Spanish and Thai.
Good reading and sources
There is no English Wikipedia article on Llama-3.1-Nemotron-Ultra-253B yet, so the facts on this page come from the sources listed above: huggingface.co/nvidia/Llama-3_1-Nemotron-Ultra-253B-v1. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.