How much memory does Nemotron-3-Super-120B-A12B need?

About 72 GB at 4-bit, 134 GB at 8-bit and 249 GB at full 16-bit precision, including an 8,192-token KV cache and about 1.5 GB of runtime overhead. These are estimates from its parameter count.

Can Nemotron-3-Super-120B-A12B run on a 24 GB graphics card?

Not entirely: even at 4-bit it needs about 72 GB, so part of it would sit in system memory, which is much slower.

What is Nemotron-3-Super-120B-A12B’s context window?

Its configuration sets 262,144 tokens. Longer contexts need more memory for the KV cache, so the practical limit on a given machine can be lower.

ANVE: the private AI office, on the record. See what that means

ANVE

Run it on your own hardware · updated 2026-10-16

Nemotron-3-Super-120B-A12B
The memory it needs, what you may do with it, and the hardware it suits.

Short answer. Nemotron-3-Super-120B-A12B needs about 72 GB of memory at 4-bit, 134 GB at 8-bit and 249 GB at full 16-bit precision, including runtime overhead and an 8,192-token KV cache. It has about 123.6 billion parameters and a 262,144-token context window. Licence in short: custom licence: check before commercial use. By memory alone, on ANVE’s launch configurations: ANVE One does not fit, ANVE Office only with partial offload to system memory, which is much slower, ANVE Office Pro (GPU) only with partial offload to system memory, which is much slower. These are estimates from its published size, not benchmarks.

What it is

Nemotron-3-Super-120B-A12B is a language model for chat, drafting and question answering, published by NVIDIA on Hugging Face on 10 March 2026. It has about 123.6 billion parameters and a configured context window of 262,144 tokens. It is a mixture-of-experts model: 512 experts, 22 used for each token, so only about 12B parameters are active at a time.

Its card lists the licence as a custom licence (nvidia-nemotron-open-model-license). In short, the licence sets its own terms; read it before any commercial use.

Hugging Face counted about 1,057,809 downloads of it in the 30 days before we checked (7 Oct 2026), and 429 people have liked it there.

The card lists 7 languages: en, fr, es, it, de, ja, zh (ISO codes as given on the card).

How much memory it needs

We size it the way you would size any model: the weights (parameters × bits ÷ 8), the KV cache for 8,192 tokens of context, and roughly 1.5 GB for the runtime itself. Real use varies with settings.

Each token of context adds about 0.09 MB of KV cache at 16-bit, so 32,768 tokens would add about 3.0 GB.

Precision	Weights	KV cache (8,192 tokens)	Total, estimated
BF16 (full precision)	247 GB	0.74 GB	249 GB
8-bit (Q8_0)	131 GB	0.74 GB	134 GB
4-bit (Q4_K)	70 GB	0.74 GB	72 GB

Architecture at a glance

Setting	Value in config.json
Architecture class	NemotronHForCausalLM
Model type	nemotron_h
Layers	88
Hidden size	4,096
Attention heads	32
Key-value heads	2 (grouped-query attention)
Head size	128
Vocabulary	131,072
Experts	512, 22 active per token
Configured context	262,144
Stored precision	bfloat16

Will it fit on ANVE?

These verdicts are arithmetic, not tests, against the specification at launch: ANVE One can give a model about three quarters of its 64 GB of unified memory, while ANVE Office and ANVE Office Pro hold it on a 24 GB graphics card and keep the rest of the stack in 64 GB or 128 GB of system memory. ANVE’s tested default is Qwen 3.8 27B; this model is not yet validated on ANVE.

Configuration (at launch)	Memory for the model	Verdict
ANVE One	about 48 GB of 64 GB unified memory	does not fit
ANVE Office	24 GB on the RTX PRO 4000 Blackwell SFF Edition	only with partial offload to system memory, which is much slower
ANVE Office Pro (GPU)	24 GB on the RTX PRO 4000 Blackwell	only with partial offload to system memory, which is much slower

What an office would use it for

Models this large are stronger on long, hard drafting and reasoning, but need most of a 64 to 128 GB machine even at 4-bit, which leaves less room for long documents and several users.

Similar models to compare

gpt-oss-120b (116.8B parameters, about 67 GB at 4-bit before context; commercial use allowed)
GLM-4.5-Air (110.5B parameters, about 64 GB at 4-bit before context; commercial use allowed)
sarvam-105b (106.0B parameters, about 61 GB at 4-bit before context; commercial use allowed)

Questions people ask

How much memory does Nemotron-3-Super-120B-A12B need?: About 72 GB at 4-bit, 134 GB at 8-bit and 249 GB at full 16-bit precision, including an 8,192-token KV cache and about 1.5 GB of runtime overhead. These are estimates from its parameter count.
Can Nemotron-3-Super-120B-A12B run on a 24 GB graphics card?: Not entirely: even at 4-bit it needs about 72 GB, so part of it would sit in system memory, which is much slower.
Is Nemotron-3-Super-120B-A12B free for commercial use?: Its card lists a custom licence (nvidia-nemotron-open-model-license), which sets its own terms; read it before any commercial use. Read the licence itself before relying on this summary; it is not legal advice.
What is Nemotron-3-Super-120B-A12B’s context window?: Its configuration sets 262,144 tokens. Longer contexts need more memory for the KV cache, so the practical limit on a given machine can be lower.

What is not real yet

No sealed ANVE unit exists; every image of the appliance on this site is a labelled render. The software runs today on a development machine. No customers yet, no certifications, no delivery date, and the final price is not announced.

Reservations are ₹999 in India or US$12 elsewhere, refundable in full until you confirm an order in writing (terms). Offices that cannot use cloud AI can apply to be one of ten design partners.

Nemotron-3-Super-120B-A12BThe memory it needs, what you may do with it, and the hardware it suits.