---
license: other
tags:
- heal
- horizon
---

# Qwen3-VL-4B-Instruct

**Original model repository:** [Qwen/Qwen3-VL-4B-Instruct](https://huggingface.co/Qwen/Qwen3-VL-4B-Instruct)

## Model Introduction

Qwen3-VL-4B-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 | 4.438B |
| Vision model (ViT) parameters | 415.3M |
| Language model (LM) parameters | 4.022B |

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

### Performance Metrics

| Chips | Data Type | ViT Image Size | Sequence Length (tokens) | Maximum Context Length (tokens) | BPU Cores (ViT / Prefill / Decode) | ViT Latency (ms) | TTFT (ms) | Prefill TPS (token/s) | Decode TPS (token/s) | BPU Memory (GB) | CPU Memory (GB) |
|---|---|---:|---:|---:|---:|---:|---:|---:|---:|---:|---:|
| J6P | W8A8 | 448 × 448 | 512 | 1024 | 4 / 4 / 4 | 18.593 | 136.189 | 4,539.468 | 31.398 | 5.6 | 1.6 |
| S6P | W8A8 | 448 × 448 | 512 | 1024 | 4 / 4 / 4 | 16.343 | 117.912 | 5,255.315 | 32.677 | 5.9 | 1.6 |

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