gemma-4-26B-A4B-it

Original model repository: google/gemma-4-26B-A4B-it

Model Introduction

Gemma 4 26B A4B is an instruction-tuned multimodal Mixture-of-Experts (MoE) model developed by Google DeepMind. It supports text and image inputs and generates text output, making it suitable for text generation, reasoning, coding, visual understanding, and multimodal tasks. The model contains approximately 25.2B parameters while activating approximately 3.8B parameters during inference.

Deployment Metrics

Model Parameters

Metric	Value
Total model parameters	25.2B
Active model parameters	3.8B
Vision model (ViT) parameters	~550M
Language model (LM) parameters	~24.65B

Parameter counts follow the upstream model details. The language model parameter count excludes the vision encoder.

Performance Metrics

Chips	Scenario	Data Type	Input	Soft Tokens	KV Cache (tokens)	ViT Latency (ms)	TTFT (ms)	Prefill TPS (token/s)	Decode TPS (token/s)	BPU Memory (GB)	CPU Memory (GB)
S6P	LLM	W4A8	2K tokens	-	0	-	981	2,124.7	32.83	24.3	3.2
S6P	LLM	W4A8	5K tokens	-	10K	-	3,022	2,001.3	31.18	24.3	3.2
S6P	LLM	W4A8	10K tokens	-	40K	-	7,052	1,450.1	28.50	24.3	3.2
S6P	Image	W4A8	4 images	266	1K	612	1,612	2,224.7	33.0	25.9	3.2
S6P	Image	W4A8	8 images	266	1K	1,230	3,087	2,338.7	32.8	25.9	3.2
S6P	Image	W4A8	16 images	266	1K	2,447	5,259	2,244.3	32.6	25.9	3.2
S6P	Image	W4A8	32 images	266	1K	4,902	9,863	2,141.6	32.0	25.9	3.2
S6P	Video	W4A8	60 images	64	1K	1,942	4,869	2,298.1	32.3	24.1	3.2

Note: Performance was measured on an S6P AIPC device. TTFT includes preprocessing and ViT latency for VLM scenarios. Memory values represent the peak memory usage measured during the specified performance test.