Qwen 2.5-VL 72B
Qwen 2.5 vision-language flagship at 72B. Strong on document understanding + multi-image queries. Qwen License.
Our verdict
Positioning
Qwen 2.5-VL 72B is the largest model in Alibaba's Qwen2.5-VL series, combining the Qwen 2.5 72B text backbone with a dedicated vision encoder for image, document, and chart understanding. It targets operators who want strong visual understanding with Qwen's multilingual text capability under the Qwen License. The model competes directly with InternVL 2.5 78B and Llama 3.2 90B Vision for the open-weight VL crown.
Strengths
- Qwen License. Commercial use is allowed, but a product with more than 100 million monthly active users must request a license from Alibaba Cloud.
- Strong Chinese + English visual understanding. Qwen's training pipeline is optimized for Chinese visual content (documents, signage, web content), giving it an edge on Chinese-language visual tasks that other VL models lack.
- Document and chart understanding capability. Qwen 2.5-VL performs well on DocVQA, ChartQA, and document-oriented visual tasks — competitive with dedicated document models.
- 72B text backbone with Qwen 2.5 generation quality. The text reasoning underlying visual tasks benefits from Qwen 2.5's strong multilingual and reasoning capabilities. Text-in-image understanding (OCR, document reading) is competitive.
- Dynamic resolution processing. Can adapt to different image resolutions for optimal detail extraction vs processing speed — important for mixed document types.
- Fits workstation hardware. Ollama's Q4_K_M build, including the vision encoder, is 48.7 GB and fits Mac Studio M3 Ultra. A single RTX 5090 cannot hold the weights: the maker's Q3_K_M files for the text-only Qwen2.5-72B-Instruct alone total 35.5 GB.
Limitations
- Text-only performance trails dedicated text models. While a 72B text backbone is capable, it was optimized for VL alignment — text-only tasks (coding, general reasoning) are below the Qwen 2.5 72B Instruct text-only model.
- Vision encoder adds memory overhead. It is about 0.7B parameters, roughly 1.4 GB at FP16, and each image adds visual tokens to the KV cache (4 to 16,384 per image by default, per the model card).
- Ecosystem support trails Llama Vision. While vLLM and SGLang support Qwen VL architectures, the community tooling and deployment examples are thinner than for Meta's Llama Vision models.
Compared with similar models
- vs InternVL 2.5 78B: Both top-tier Chinese-lab VL models. InternVL has stronger general visual benchmarks; Qwen 2.5-VL may have stronger Chinese visual content performance. Both are excellent — pick by specific visual benchmark priorities and licensing.
- vs Llama 3.2 90B Vision Instruct: Qwen 2.5-VL has stronger Chinese visual performance; Llama 3.2 90B Vision has larger text backbone and broader ecosystem. Pick by license + language needs.
- vs Qwen 2.5 72B Instruct (text-only): Qwen 2.5-VL adds vision to a Qwen 2.5 generation (not 3) text backbone. Slightly older text quality but adds visual capability. Pick by modality need — text-only operators should pick Qwen 2.5 72B Instruct.
Should you run this locally?
Yes if you need visual understanding at the 70B scale, have Chinese-language visual content in your workload, and have workstation-class hardware. The combination of Qwen text quality and visual capability is the draw in the 70B VL class.
No if your workload is text-only (pick Qwen 2.5 72B Instruct), you prioritize maximum visual benchmark performance (compare InternVL 2.5 78B), or you need Llama ecosystem compatibility (pick Llama Vision models).
How it compares
- vs InternVL 2.5 78B: Both top VL models. InternVL stronger general visual; Qwen potentially wins on Chinese visual.
- vs Llama 3.2 90B Vision Instruct: Llama wins text backbone size and ecosystem.
- vs Qwen 2.5 72B Instruct: VL vs text-only. Pick by modality need.
- vs Molmo 72B: Qwen has broader visual scope; Molmo has pointing/grounding capabilities.
Run this yourself
- Datacenter: Single A100 80GB at FP8 with vLLM vision support.
- Workstation: Mac Studio M3 Ultra (96 GB or 256 GB) at Q4.
- Consumer: A single RTX 5090 (32 GB) needs partial offload to system RAM even at Q3.
- Vendor: huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct
Overview
Qwen 2.5 vision-language flagship at 72B. Strong on document understanding + multi-image queries. Qwen License.
How to run it
Qwen 2.5 VL 72B is Alibaba's 72B vision-language model — 72B dense text backbone + Qwen-VL vision encoder. Run at Q4_K_M via llama.cpp with llama-server and an mmproj file for vision, or Ollama with the qwen2.5vl:72b tag. Ollama's Q4_K_M file, including the vision encoder, is 48.7 GB. Minimum VRAM: 48 GB — RTX A6000 at Q3_K_M with vision, or text-only Q4_K_M. Recommended: A100 80GB at AWQ-INT4 for full vision serving. Qwen-VL's vision encoder is well-optimized — lower VRAM overhead than InternVL's InternViT. Strong OCR, document understanding, and visual reasoning. The model card says config.json is set for context up to 32,768 tokens, though the repo's config.json lists max_position_embeddings 128000. The card describes YaRN for longer inputs but does not recommend it. Qwen 2.5 VL uses the same text backbone as Qwen 2.5 72B — broad ecosystem support. For production: vLLM with multimodal pipeline and tensor-parallel if needed.
Hardware guidance
Minimum: RTX A6000 48GB at Q3_K_M + vision. Recommended: A100 80GB at AWQ-INT4. VRAM math: Ollama's Q4_K_M file, including the vision encoder, is 48.7 GB. KV cache at 8K: ~2.7 GB. Total: ~51.4 GB. A6000 48GB: tight at Q4_K_M + vision + 8K, while Q3_K_M (text weights about 35.5 GB) leaves more room. A100 80GB: comfortable for Q4 + vision + 8K. Dual RTX 4090 48 GB: Q4_K_M text-only, or Q3_K_M + vision. Mac Studio M3 Ultra 96GB: Q4_K_M + vision. Qwen-VL's encoder is smaller than InternViT — better VRAM efficiency for vision. The maker's AWQ-INT4 build is 43 GB, enabling 16K+ context on 80 GB.
What breaks first
- Vision tag in Ollama. The official Ollama tag is
qwen2.5vl:72b. Test with raw llama.cpp if Ollama fails. 2. Image preprocessing mismatch. Qwen-VL expects specific image preprocessing (resolution, normalization). Feeding raw images without preprocessing degrades vision quality. Use the model's image processor from HF transformers. 3. KV cache with vision. Vision tokens are prepended to the text prompt — each image adds 4 to 16,384 tokens by default, or 256 to 1,280 with the model card's example min_pixels and max_pixels. Multiple images inflate context and KV cache proportionally. Budget for image tokens. 4. Qwen 2.5 vs Qwen 3 VL. Qwen 2.5 VL and Qwen 3 VL use different vision encoders. Don't mix model files or chat templates between versions.
Runtime recommendation
Common beginner mistakes
Mistake: Using a standard text-only GGUF and expecting vision to work. Fix: With llama.cpp, vision requires the separate mmproj GGUF alongside the text GGUF, such as ggml-org/Qwen2.5-VL-72B-Instruct-GGUF. Mistake: Ignoring image token count in context budget. Fix: Each image in Qwen-VL consumes 4 to 16,384 tokens by default, or 256 to 1,280 with the model card's example settings. Subtract image tokens from your available context window. Mistake: Using Llama 3.2 Vision chat template for Qwen-VL. Fix: Different architectures, different templates. Use Qwen's chat template from tokenizer_config.json. Mistake: Sending large images without preprocessing. Fix: Qwen-VL expects images within a specific resolution range. Use Qwen's image processor or resize manually before inference.
Family & lineage
How this model relates to others in its lineage. Family members share architecture and training-data roots; parent / children edges record direct distillation or fine-tune relationships.
Strengths
- Frontier-tier multimodal
- Strong document Q&A
Weaknesses
- 48GB+ VRAM tier
Quantization variants
Each quantization trades model quality for file size and VRAM. Q4_K_M is the most popular starting point.
| Quantization | File size | VRAM required |
|---|---|---|
| AWQ-INT4 | 42.0 GB | 48 GB |
At Q4 Qwen 2.5-VL 72B needs about 66 GB of memory by our estimate, more than a 32 GB graphics card holds. Renting a GPU with enough memory (or several GPUs) by the hour is the practical way to try it before you plan hardware for it.
Referral links: RunLocalAI may earn a commission or account credits if you sign up and rent, at no extra cost to you. Hourly rates and which GPUs are listed change all the time, so check the provider's page before you start. How we make money.
Get the model
HuggingFace
Original weights
Source repository — direct quantization required.
Hardware that runs this
Cards with enough VRAM for at least one quantization of Qwen 2.5-VL 72B.
Models worth comparing
Same parameter band, plus what's one tier above and below — so you can decide what actually fits your hardware.
Frequently asked
What's the minimum VRAM to run Qwen 2.5-VL 72B?
Can I use Qwen 2.5-VL 72B commercially?
What's the context length of Qwen 2.5-VL 72B?
Does Qwen 2.5-VL 72B support images?
Source: huggingface.co/Qwen/Qwen2.5-VL-72B-Instruct
Reviewed by RunLocalAI Editorial. See our editorial policy for how we research and verify model claims.
Related — keep moving
Verify Qwen 2.5-VL 72B runs on your specific hardware before committing money.