Model comparison
Qwen3-VL vs Qwen2.5-VL
- Input
- $0.20 / 1M tokens (0-32K tier)
- Output
- $1.60 / 1M tokens (0-32K tier)
- Context
- 256K
- Released
- 2025-09-23
- Input
- $0.80 per 1M input tokens
- Output
- $1.00 per 1M output tokens
- Context
- 32,768 tokens (extendable to ~64K for long video via YaRN; some API providers expose up to 131K)
- Released
- 2025-01-26
Qwen3-VL scores higher on 5 of the 5 benchmarks both models report; Qwen3-VL is cheaper on input tokens ($0.2 vs $0.8 per 1M, 4×); Qwen3-VL has the larger context window (256K vs 33K).
Specs and pricing side by side
| Spec | Qwen3-VL | Qwen2.5-VL |
|---|---|---|
| Maker | Alibaba (Qwen) | Alibaba (Qwen)same |
| Released | 2025-09-23 | 2025-01-26 |
| Context window | 256K | 32,768 tokens (extendable to ~64K for long video via YaRN; some API providers expose up to 131K) |
| Max output | 32K | Not separately documented; bounded by the… |
| Input price | $0.20 / 1M tokens (0-32K tier) | $0.80 per 1M input tokens |
| Output price | $1.60 / 1M tokens (0-32K tier) | $1.00 per 1M output tokens |
| Parameters | 235B total / 22B active (flagship MoE); also dense 2B/4B/8B/32B and MoE 30B-A3B | Four open sizes: 3B, 7B, 32B and 72B (each in base and instruction-tuned variants) |
| License | Apache 2.0 | Mixed by size: Qwen2.5-VL-72B-Instruct under the Qwen License; 7B and 32B under Apache 2.0; 3B under the Qwen Research License (non-commercial only). |
| Weights | Open weights | Open weightssame |
| Modalities | Text, Vision, Video, PDF | text, image, video |
| Knowledge cutoff | Not publicly disclosed | Not officially published by Alibaba |
Benchmarks both models report
| Benchmark | Qwen3-VL | Qwen2.5-VL |
|---|---|---|
| MathVista | 85.8% | 74.8% |
| MathVision | 74.6% | 38.1% |
| DocVQA | 96.5% | 96.4% |
| MMMU-Pro | 69.3% | 51.1% |
| ScreenSpot Pro | 61.8% | 43.6% |
Scores are as published by each maker or its cited source; harness, reasoning effort and date can differ between rows, so treat small gaps as ties. Sources are listed on each model page.
Which one should you pick?
Pick Qwen3-VL if…
- You care about MathVista, MathVision, DocVQA (it scores higher on 5 shared benchmarks).
- Cost matters: input tokens cost 4× less ($0.2 vs $0.8 per 1M).
- You need the longer context (256K vs 33K).
- You need Text / Vision / Video / PDF input, which Qwen2.5-VL lacks.
- Your use case is: document, invoice, and form parsing with multilingual OCR.
Pick Qwen2.5-VL if…
- You need text / image / video input, which Qwen3-VL lacks.
- Your use case is: document AI: extracting text, tables and layout from invoices, contracts, research papers and scanned forms.
FAQ
Which is better, Qwen3-VL or Qwen2.5-VL?
Qwen3-VL scores higher on 5 of the 5 benchmarks both models report; Qwen3-VL is cheaper on input tokens ($0.2 vs $0.8 per 1M, 4×); Qwen3-VL has the larger context window (256K vs 33K). The better pick depends on whether you value benchmark scores, price per token, context length or openness for your workload.
How much cheaper is Qwen3-VL?
Qwen3-VL: $0.20 / 1M tokens (0-32K tier) input, $1.60 / 1M tokens (0-32K tier) output. Qwen2.5-VL: $0.80 per 1M input tokens input, $1.00 per 1M output tokens output (list API prices as published by each maker).
What are the context windows of Qwen3-VL and Qwen2.5-VL?
Qwen3-VL offers 256K with up to 32K output; Qwen2.5-VL offers 32,768 tokens (extendable to ~64K for long video via YaRN; some API providers expose up to 131K) with up to Not separately documented; bounded by the 32,768-token context window output.
Is Qwen3-VL newer than Qwen2.5-VL?
Qwen3-VL was released 2025-09-23; Qwen2.5-VL was released 2025-01-26.