Qwen2-VL Models by Alibaba
Last refreshed 2026-05-19. Next refresh: weekly.
Details
Capabilities
About
Qwen2-VL is Alibaba’s Qwen Team vision–language series built on Qwen2, the successor to Qwen-VL. The family is aimed at images of many resolutions and aspect ratios, long video understanding, visual agents that can drive phones or robots from screen context, and multilingual text in images beyond English and Chinese. Open sizes include 2B and 7B Instruct models, with a larger 72B route; architecture notes highlight Naive Dynamic Resolution visual tokens and Multimodal Rotary Position Embedding (M-RoPE) for text, 2D image, and video position.
Decision facts
- Best fit
- vision and multimodal workagent workflows
- Capability starting point
- Qwen2-VL-72B-Instruct with 32K context and multimodal inputs
- Lowest tracked input
- Qwen2-VL-72B-Instruct · $0.900/1M · Fireworks AI
- Closest related family
- Ovis Image
Current Variants
Use-when guidance is based on each model's tracked capabilities, context window, release date, and replacement status.
Use when the workload needs multimodal, 32K context, and 72B parameters.
| Model | Use when | Released | Signals | Status |
|---|---|---|---|---|
| Qwen2-VL-72B-Instruct | Use when the workload needs multimodal, 32K context, and 72B parameters. | 2025-01 | multimodal32K context72B parameters | Current |
Release Timeline
1 release groupSpecifications(1 models)
| Model | Released | Context | Parameters | Vision | Multimodal |
|---|---|---|---|---|---|
| Qwen2-VL-72B-Instruct | 2025-01 | 32K | 72B | Yes | Yes |
Available From(1 provider)
Pricing
| Model | Provider | Input / 1M | Output / 1M | Type |
|---|---|---|---|---|
| Qwen2-VL-72B-Instruct | Fireworks AI | $0.9 | $0.9 | Serverless |
