Qwen2.5-VL 72B Instruct
Summary
Qwen2.5-VL 72B Instruct is an open-weight language model from Alibaba (Qwen), released on 26 Jan 2025. It accepts text and images and generates text, with a 128K-token context window and up to 115K output tokens per response. As of 18 Nov 2026, the lowest listed API price is $0.80 per million input tokens and $1 per million output tokens (OpenRouter), across 1 provider we track.
Qwen vision-language model for visual reasoning, documents, and agent tasks
Last verified 18 Nov 2026. Every figure on this page links to its source.
Specifications
| Developer | Alibaba (Qwen) |
|---|---|
| Release date | 26 Jan 2025 [models.dev] |
| Status | Generally available [models.dev] |
| Weights | Open weights [Hugging Face] |
| License | qwen [Hugging Face] |
| Parameters | 73.4B [Hugging Face] |
| Context window | 128K (128,000 tokens) [models.dev] |
| Max output | 115K tokens [models.dev] |
| Knowledge cutoff | Apr 2024 [models.dev] |
| Input | Text and images [models.dev] |
| Output | Text [models.dev] |
| Reasoning mode | No [models.dev] |
| Tool calling | No [models.dev] |
| Structured output | Yes [models.dev] |
| Hugging Face | Qwen/Qwen2.5-VL-72B-Instruct [Hugging Face] |
Qwen2.5-VL 72B Instruct API pricing by provider
USD per million tokens. Sorted by input price.
| Provider | Input | Output | Cached input | Context | Source |
|---|---|---|---|---|---|
| OpenRouter qwen/qwen2.5-vl-72b-instruct | $0.80 | $1 | $0.40 | 128K | OpenRouter, 18 Nov 2026 |
Running Qwen2.5-VL 72B Instruct locally
Estimated memory, from the parameter count and the architecture in config.json on Hugging Face. Includes about 0.8 GB of runtime overhead.
| Quantization | Weights | Total, 8K context | Total, 32K context |
|---|---|---|---|
| Q4_K_M (4-bit) | 45 GB | 48 GB | 56 GB |
| Q8_0 (8-bit) | 78 GB | 81 GB | 90 GB |
| FP16 / BF16 (16-bit) | 147 GB | 150 GB | 158 GB |
At 4-bit with an 8K conversation, Qwen2.5-VL 72B Instruct needs about 48 GB, more than any single consumer graphics card; it needs a unified-memory machine, a multi-GPU setup or a split between graphics card and system RAM. Check your own hardware, or see Best computer for local AI on our sister site AIPCs.
Frequently asked questions
What is the context window of Qwen2.5-VL 72B Instruct?
Qwen2.5-VL 72B Instruct has a context window of 128,000 tokens (128K) and can generate up to 115,200 tokens in a single response, according to models.dev.
How much does the Qwen2.5-VL 72B Instruct API cost?
Qwen2.5-VL 72B Instruct is listed from $0.80 per million input tokens and $1 per million output tokens (OpenRouter). Prices last verified 18 Nov 2026.
Is Qwen2.5-VL 72B Instruct open source?
Qwen2.5-VL 72B Instruct is an open-weight model: its weights are published on Hugging Face as Qwen/Qwen2.5-VL-72B-Instruct under the qwen license. Check the license terms before commercial use.
When was Qwen2.5-VL 72B Instruct released?
Alibaba (Qwen) released Qwen2.5-VL 72B Instruct on 26 Jan 2025, according to models.dev.
How much VRAM does Qwen2.5-VL 72B Instruct need?
About 48 GB at 4-bit (Q4_K_M) with an 8K-token conversation, and about 81 GB at 8-bit. That is more than any single consumer graphics card holds. These are estimates from the model's published parameter count and architecture.
Similar models from other labs
Comparable on output type, weights, release date, price and context window. Prices are the lowest listed, USD per million tokens.
| Model | Lab | Released | Context | Input $/1M | Output $/1M | Weights |
|---|---|---|---|---|---|---|
| Aya Expanse 32B | Cohere | 24 Oct 2024 | 128K | $0.50 | $1.50 | Open |
| Llama-4-Maverick-17B-128E-Instruct-FP8 | Meta | 5 Apr 2025 | 128K | $0.24 | $0.85 | Open |
| DeepSeek V3 0324 | DeepSeek | 24 Mar 2025 | 164K | $0.25 | $1 | Open |
| Llama 3.1 Euryale 70B v2.2 | Sao10k | 28 Aug 2024 | 131K | $0.85 | $0.85 | Open |
| DeepSeek-V3 | DeepSeek | 26 Dec 2024 | 64K | $0.32 | $0.89 | Open |
| Llama 3.1 70B Instruct | Meta | 23 Jul 2024 | 128K | $0.72 | $0.72 | Open |
Change history
- Structured output changed from none to yes
- Release date changed from Sept 2024 to 26 Jan 2025
- Tool calling changed from yes to no
- Status changed from deprecated to active
- Context window changed from 131K to 128K
- Max output changed from 8K to 115K
- No longer listed on alibaba
- Status changed from active to deprecated