How much VRAM does Gemma 4 26B A4B need?

51.6 GB for its weights at 16-bit, 25.8 GB at 8-bit and 14.5 GB at 4-bit, plus the KV cache: 20.5 KB for each token, 881 MB for a conversation of 32K tokens and 5.58 GB at its longest context of 262,144 tokens. The server also keeps about 8% of each GPU free.

Can Gemma 4 26B A4B run on 1 GPU?

Yes, at 4-bit: the smallest single GPU here that holds it with a 32K conversation is the L4.

How many H100s does Gemma 4 26B A4B need?

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

Can I run Gemma 4 26B A4B on an RTX 5090?

The RTX 5090 has 32 GB. With a 32K conversation, Gemma 4 26B A4B needs 1 GeForce RTX 5090 at 4-bit and 2 GeForce RTX 5090s at 16-bit.

How fast does Gemma 4 26B A4B run?

At 4-bit on 1 H100, memory lets 1 conversation get at most 1,070 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 Gemma 4 26B A4B need?

Gemma 4 26B A4B by Google has 25.8B parameters, 4B of them active for each token. Its weights take 51.6 GB at 16-bit, 25.8 GB at 8-bit and 14.5 GB at 4-bit; released in BF16, its files are 51.6 GB. Each token of a conversation adds 20.5 KB of KV cache, so a 32K conversation holds 881 MB.

Size Gemma 4 26B A4B for your setup Questions

Google · released in BF16

Gemma 4 26B A4B at a glance

Parameters: 25.8B; 4B active for each token
As released: 51.6 GB; Its BF16 files
At 16-bit: 51.6 GB
At 4-bit: 14.5 GB; NVFP4
KV cache a token: 20.5 KB; At 16-bit
Longest context: 262,144; 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	4 GPUs135 tokens/s	2 GPUs123 tokens/s	1 GPU96 tokens/s
GeForce RTX 5090	32 GB	2 GPUs404 tokens/s	1 GPU367 tokens/s	1 GPU572 tokens/s
L40S	48 GB	2 GPUs195 tokens/s	1 GPU177 tokens/s	1 GPU276 tokens/s
A100 80GB	80 GB	1 GPU230 tokens/s	1 GPU418 tokens/s	1 GPU651 tokens/s
H100	80 GB	1 GPU377 tokens/s	1 GPU686 tokens/s	1 GPU1,070 tokens/s
RTX PRO 6000 Blackwell	96 GB	1 GPU202 tokens/s	1 GPU367 tokens/s	1 GPU572 tokens/s
DGX Spark	128 GB	1 GPU31 tokens/s	1 GPU56 tokens/s	1 GPU87 tokens/s
H200	141 GB	1 GPU540 tokens/s	1 GPU983 tokens/s	1 GPU1,533 tokens/s
B200	180 GB	1 GPU901 tokens/s	1 GPU1,639 tokens/s	1 GPU2,555 tokens/s
Instinct MI300X	192 GB	1 GPU597 tokens/s	1 GPU1,086 tokens/s	1 GPU1,693 tokens/s
Instinct MI325X	256 GB	1 GPU676 tokens/s	1 GPU1,229 tokens/s	1 GPU1,916 tokens/s
Instinct MI355X	288 GB	1 GPU901 tokens/s	1 GPU1,639 tokens/s	1 GPU2,555 tokens/s

Models of a similar size

Every model

Questions

How much VRAM does Gemma 4 26B A4B need?: 51.6 GB for its weights at 16-bit, 25.8 GB at 8-bit and 14.5 GB at 4-bit, plus the KV cache: 20.5 KB for each token, 881 MB for a conversation of 32K tokens and 5.58 GB at its longest context of 262,144 tokens. The server also keeps about 8% of each GPU free.
Can Gemma 4 26B A4B run on 1 GPU?: Yes, at 4-bit: the smallest single GPU here that holds it with a 32K conversation is the L4.
How many H100s does Gemma 4 26B A4B need?: With 1 conversation of 32K tokens, 1 H100 at 16-bit and 1 H100 at 4-bit. Each H100 has 80 GB, of which vLLM uses 92% by default.
Can I run Gemma 4 26B A4B on an RTX 5090?: The RTX 5090 has 32 GB. With a 32K conversation, Gemma 4 26B A4B needs 1 GeForce RTX 5090 at 4-bit and 2 GeForce RTX 5090s at 16-bit.
How fast does Gemma 4 26B A4B run?: At 4-bit on 1 H100, memory lets 1 conversation get at most 1,070 tokens a second. Real servers reach part of that ceiling, and each conversation gets less when many run at once.

Sources

Gemma 4 26B A4B at a glance

GPUs needed at each number format

Models of a similar size

Questions

Where each fact comes from