How much VRAM does Llama 4 Scout need?

217 GB for its weights at 16-bit, 109 GB at 8-bit and 61.1 GB at 4-bit, plus the KV cache: 49.2 KB for each token, 2.82 GB for a conversation of 32K tokens and 517 GB at its longest context of 10,485,760 tokens. The server also keeps about 8% of each GPU free.

Can Llama 4 Scout run on 1 GPU?

Yes, at 4-bit: the smallest single GPU here that holds it with a 32K conversation is the A100 80GB.

How many H100s does Llama 4 Scout need?

With 1 conversation of 32K tokens, 4 H100s at 16-bit and 1 H100 at 4-bit. Each H100 has 80 GB, of which vLLM uses 92% by default.

Can I run Llama 4 Scout on an RTX 5090?

The RTX 5090 has 32 GB. With a 32K conversation, Llama 4 Scout needs 4 GeForce RTX 5090s at 4-bit and 8 GeForce RTX 5090s at 16-bit.

How fast does Llama 4 Scout run?

At 4-bit on 1 H100, memory lets 1 conversation get at most 271 tokens a second. Real servers reach part of that ceiling, and each conversation gets less when many run at once.

How much GPU memory does Llama 4 Scout need?

Llama 4 Scout by Meta has 109B parameters, 17B of them active for each token. Its weights take 217 GB at 16-bit, 109 GB at 8-bit and 61.1 GB at 4-bit; released in BF16, its files are 217 GB. Each token of a conversation adds 49.2 KB of KV cache, so a 32K conversation holds 2.82 GB.

Size Llama 4 Scout for your setup Questions

Meta · released in BF16

Llama 4 Scout at a glance

Parameters: 109B; 17B active for each token
As released: 217 GB; Its BF16 files
At 16-bit: 217 GB
At 4-bit: 61.1 GB; NVFP4
KV cache a token: 49.2 KB; At 16-bit
Longest context: 10,485,760; Tokens

GPUs needed at each number format

For 1 conversation of 32,768 tokens with a 16-bit KV cache, counting 92% of each GPU's memory as vLLM does. Under each count, the most tokens a second memory allows.

GPU	Memory	16-bit	8-bit	4-bit
L4	24 GB	16 GPUs65 tokens/s	8 GPUs121 tokens/s	4 GPUs97 tokens/s
GeForce RTX 5090	32 GB	8 GPUs389 tokens/s	4 GPUs362 tokens/s	4 GPUs579 tokens/s
L40S	48 GB	8 GPUs188 tokens/s	4 GPUs174 tokens/s	2 GPUs140 tokens/s
A100 80GB	80 GB	4 GPUs222 tokens/s	2 GPUs206 tokens/s	1 GPU165 tokens/s
H100	80 GB	4 GPUs364 tokens/s	2 GPUs338 tokens/s	1 GPU271 tokens/s
RTX PRO 6000 Blackwell	96 GB	4 GPUs195 tokens/s	2 GPUs181 tokens/s	1 GPU145 tokens/s
DGX Spark	128 GB	2 GPUs15 tokens/s	1 GPU14 tokens/s	1 GPU22 tokens/s
H200	141 GB	2 GPUs261 tokens/s	1 GPU242 tokens/s	1 GPU388 tokens/s
B200	180 GB	2 GPUs435 tokens/s	1 GPU404 tokens/s	1 GPU646 tokens/s
Instinct MI300X	192 GB	2 GPUs288 tokens/s	1 GPU267 tokens/s	1 GPU428 tokens/s
Instinct MI325X	256 GB	1 GPU163 tokens/s	1 GPU303 tokens/s	1 GPU485 tokens/s
Instinct MI355X	288 GB	1 GPU217 tokens/s	1 GPU404 tokens/s	1 GPU646 tokens/s

Models of a similar size

Every model

Questions

How much VRAM does Llama 4 Scout need?: 217 GB for its weights at 16-bit, 109 GB at 8-bit and 61.1 GB at 4-bit, plus the KV cache: 49.2 KB for each token, 2.82 GB for a conversation of 32K tokens and 517 GB at its longest context of 10,485,760 tokens. The server also keeps about 8% of each GPU free.
Can Llama 4 Scout run on 1 GPU?: Yes, at 4-bit: the smallest single GPU here that holds it with a 32K conversation is the A100 80GB.
How many H100s does Llama 4 Scout need?: With 1 conversation of 32K tokens, 4 H100s at 16-bit and 1 H100 at 4-bit. Each H100 has 80 GB, of which vLLM uses 92% by default.
Can I run Llama 4 Scout on an RTX 5090?: The RTX 5090 has 32 GB. With a 32K conversation, Llama 4 Scout needs 4 GeForce RTX 5090s at 4-bit and 8 GeForce RTX 5090s at 16-bit.
How fast does Llama 4 Scout run?: At 4-bit on 1 H100, memory lets 1 conversation get at most 271 tokens a second. Real servers reach part of that ceiling, and each conversation gets less when many run at once.

Sources

Llama 4 Scout at a glance

GPUs needed at each number format

Models of a similar size

Questions

Where each fact comes from