QVQ by Qwen
by Qwen team at Alibaba Cloud
Open-weight image-text reasoning model, now legacy.
Score
Score
Our verdict
QVQ by Qwen is an open-weight, 72B-parameter visual reasoning research preview built on Qwen2‑VL‑72B. It accepts image and text inputs and returns text-only outputs, with a strict single-round limit and no video support. The model posted strong research benchmarks (e.g., MMMU 70.3, MathVista-mini 71.4) but comes with vendor-documented failure modes, including language mixing, recursive reasoning loops, loss of image focus during long chains, and a need for enhanced safety. There is no subscription plan; weights are free under the custom Qwen licence, but full-precision inference demands roughly 145 GB+ VRAM and multi-GPU servers. Hosted visual reasoning access today is via Alibaba Cloud Model Studio on newer models (QVQ-Max, Qwen3-VL) billed per token. As of 2026, QVQ-72B-Preview is clearly legacy—useful as a research baseline or for controlled experiments, not for production conversational assistants or video tasks.
Overview
Score breakdown
Overall score
Scores are editorial assessments by the Compare Best AI team on a 0–10 scale.
Expert review
CBAI Editorial Team
Compare Best AI · Editorial Team
## Overview
How we tested
Days tested
7 days
Tasks evaluated
- ·Solve multi-step visual math problems from MathVista-mini
- ·Answer diagram-based science questions with chain-of-thought prompts
- ·Stress-test long reasoning chains for focus drift and loops
- ·Quantized (8-bit) vs bf16 inference latency/quality comparison
Method
Compared against QVQ-Max and Qwen3-VL (Instruct/Thinking) using identical inputs
Reviewer
CBAI Editorial Team
Plans & pricing
Self-host
Local or self-hosted use
- No public self-serve SaaS price table on the official page
Pricing may vary by region. Always verify on the vendor's website.
Feature comparison
| Feature | QVQ by Qwen | QVQ-Max | Qwen3-VL (Instruct/Thinking) |
|---|---|---|---|
| Status | |||
| Released 2024-12-25 | |||
| Released 2025-03-28 or later | |||
| Current as of 2026-08 | |||
| Access | |||
| Available via Alibaba Cloud Model Studio | |||
| I/O | |||
| Single-round dialogue only | |||
| No video input supported | |||
Is it right for you?
Good fit for
AI researchers
Study long-chain visual reasoning behavior, limitations, and failure modes in a large open-weight VL model.
Benchmarkers
Reproduce and compare MMMU, MathVista, MathVision, and OlympiadBench results on a known research baseline.
GPU-rich R&D teams
Prototype self-hosted multimodal reasoning with control over inference stack (Transformers/vLLM) on multi-GPU servers.
Model engineers
Experiment with quantization (AWQ/GPTQ/MLX 8-bit) and serving performance trade-offs.
Less suited for
Production conversational apps
Single-turn only, research-preview stability, and safety caveats make it unsuitable for production chat.
Video understanding
The model accepts images and text only; video input is not supported.
Teams without datacenter GPUs
Full-precision inference needs ~145 GB+ VRAM and multi-GPU hardware; hosted access is per-token on other models.
User reviews
Editorial score
Distribution is estimated from our editorial score. Verified user reviews coming soon.
Use cases
Typical ways teams rely on this tool — from everyday tasks to specialized workflows.
- Visual math reasoning
- Scientific diagram question answering
- Benchmark reproduction (MMMU, MathVista, MathVision)
- Olympiad-style problem analysis
- Long-chain visual reasoning studies
Integrations
Reported connectors
Apps and services commonly connected out of the box or via official connectors.
- Hugging Face Transformers
- vLLM
- ModelScope
Details
Category
Price
- Free to self-host / open weights
- hosted usage billed by the cloud vendor if used
Free version
Best for
- Academic benchmark replication and analysis
- Olympiad-style visual math/science problem solving
- Safety and hallucination behavior studies
- Quantization and inference optimization experiments
Frequently asked questions
QVQ by Qwen
Developer
Ready to get started?
Visit the QVQ by Qwen website to explore plans and pricing.
Was this page helpful?
Compare Best AI may earn a commission when you click links on this page. This does not influence our editorial scores or recommendations. Advertiser disclosure