Qwen3-VL-2B-Instruct

Original model repository: Qwen/Qwen3-VL-2B-Instruct

Model Introduction

Qwen3-VL-2B-Instruct is an instruction-tuned Vision-Language Model (VLM) for understanding images, videos, and text. It combines a vision encoder with a dense autoregressive language model. The model supports visual question answering, multilingual OCR, document understanding, visual grounding, spatial reasoning, video understanding, visual coding, and visual-agent tasks.

Deployment Metrics

Model Parameters

Metric	Value
Total model parameters	2.128B
Vision model (ViT) parameters	407.0M
Language model (LM) parameters	1.721B

Parameter counts are calculated from the tensors stored in the upstream checkpoint.

Performance Metrics

Test Configuration

Metric	Matrix6P W8A8 Image	S6P W8A8 Image	S6P W8A8 Video
ViT image size	448 × 448	448 × 448	448 × 448 per frame
Sequence length	512	512	1024
Maximum context length	1024	1024	4096
BPU cores (ViT / Prefill / Decode)	4 / 4 / 4	4 / 4 / 4	4 / 4 / 4

Performance Results

Metric	Matrix6P W8A8 Image	S6P W8A8 Image	S6P W8A8 Video
ViT latency	18.489 ms	16.662 ms	80.971 ms
Time to first token (TTFT)	76.787 ms	68.262 ms	452.232 ms
Prefill throughput	9,567.264 tokens/s	10,786.589 tokens/s	7,462.987 tokens/s
Decode throughput	61.967 tokens/s	67.446 tokens/s	43.671 tokens/s

Memory Usage

Metric	Matrix6P W8A8 Image	S6P W8A8 Image	S6P W8A8 Video
BPU memory	3.0 GB	3.1 GB	4.9 GB
CPU memory	1.3 GB	1.3 GB	1.4 GB

Note: TTFT includes preprocessing and ViT latency. Memory values represent the peak memory usage measured during the specified performance test.