Best Visual Question Answering Models
Compare the best 95 visual question answering models and try 88 of them on your own images, free in the Roboflow Playground. 36 are open-weight, so you can self-host them for free under their licenses.
95 models · 36 open-weight · 88 free to try · prices synced Oct 27, 2026
Best Visual Question Answering Models 2026
The best visual question answering model on our visual reasoning benchmark right now is Gemini 4 Argon by Google, scoring 89.4% across 63 vision language models tested, followed by GPT-6 Astra at 87.2%. Updated Oct 13, 2026. See the full ranking.
Benchmark Leader
Gemini 4 Argon
Highest score on our visual reasoning benchmark, 63 vision language models tested
Open-Weight Leader
Qwen3.8 27B
Best score among the open-weight models, #26 of 63 overall on our visual reasoning benchmark
Fastest on Benchmark
Mage-VL 4B
Quickest in our visual reasoning benchmark runs: 0.4s per request on average
Lowest Measured Cost
Qwen3.7 Flash
Cheapest in our visual reasoning benchmark runs: <$0.0001 per request as measured, not list price
Benchmark scores
Ranked on our visual reasoning benchmark, 63 vision language models tested, updated Oct 13, 2026
Open-Source Visual Question Answering Models
36 models with downloadable weights you can self-host under their licenses (Custom, MIT, and Apache 2.0). 29 run live in the Playground through hosted APIs, so self-hosting is optional.
| Actions | ||||
|---|---|---|---|---|
Mistral Large 4Mistral | 1.05T total, 49B active | Custom | Oct 2026 | |
MiMo V2.6 FlashXiaomi | 309B total, 15B active | MIT | Sep 2026 | |
MiMo V2.6 ProXiaomi | 1.02T total, 42B active | MIT | Sep 2026 | |
GLM 5.3 FlashZ.ai | 320B total, 18B active | MIT | Aug 2026 | |
Qwen3.8 27BQwen | 27.78B | Apache 2.0 | Aug 2026 | |
Muse Glimmer 30BMeta | 29.6B | Apache 2.0 | Aug 2026 | |
Kimi K3Moonshot AI | 2.8T | Modified MIT | Jul 2026 | |
Gemma 4 12BGoogle | 12B | Apache 2.0 | Jun 2026 |
Visual Question Answering Models via API
59 proprietary models where the weights aren't downloadable: access is through each provider's API and billed by them. Try all of them free in the Playground.
| Actions | |||||
|---|---|---|---|---|---|
Gemini 4 ArgonGoogle NEW | $2.00 | $10.00 | 1M | Oct 2026 | |
Muse WonderMeta | $4.00 | $20.00 | 1.0M | Oct 2026 | |
Claude Haiku 5.5Anthropic | $0.10 | $0.50 | 1M | Oct 2026 | |
GPT-6.1 SolOpenAI | $2.00 | $10.00 | 1.1M | Sep 2026 | |
Claude Sonnet 5.5Anthropic | $2.00 | $10.00 | 1M | Sep 2026 | |
Claude Opus 5.5Anthropic | $4.00 | $20.00 | 1M | Sep 2026 | |
GPT-6 LunaOpenAI | $0.10 | $0.50 | 1.1M | Sep 2026 | |
GPT-6 SolOpenAI | $2.00 | $10.00 | 1.1M | Sep 2026 |
How to Choose a Visual Question Answering Model
VQA quality tracks general reasoning ability more than any other vision task: the model must read the question, find the evidence in the image, and reason to an answer. That makes the choice mostly about how hard your questions are and how expensive a wrong answer is.
Questions that chain steps (counting plus comparison, reading text then interpreting it, spatial reasoning in cluttered scenes) separate the top models from the rest quickly. If your questions are open-ended or your users type anything they want, the frontier tier is the safe default; our visual reasoning benchmark on the evals page shows how the current models actually rank on this.
When the question is fixed and simple ("is the shelf empty?", "is a person present?"), a small open VLM answers at a fraction of the cost, and often a trained classifier or detector answers it even more reliably: a fixed question is really a classification or detection task in disguise, and converting it usually improves both accuracy and unit economics.
VQA models answer fluently whether or not they are right. For production use, prefer questions whose answers can be checked (counts you can verify with a detector, text you can verify with OCR), ask for the evidence in the answer, and treat free-form answers about safety-critical content as drafts for review rather than decisions.
The bottom line: Open-ended questions need a frontier model; a fixed, repeated question is usually a classification or detection task wearing a question mark, and converting it is cheaper and more reliable.
Frequently Asked Questions About Visual Question Answering Models
On our visual reasoning benchmark (63 models tested, updated Oct 13, 2026), Gemini 4 Argon by Google currently scores highest at 89.4%. The full ranking is on our evals page. For fixed categories in production, a model fine-tuned on your own data still often wins.
Yes. 36 of the 95 visual question answering models here are open-weight (for example Qwen3.8 27B, Qwen3.6 27B, and Qwen3.5-27B), free to self-host under their licenses (Custom, MIT, and Apache 2.0).
Yes. You can run 88 of them in the Roboflow Playground for free. Upload an image and compare the models' output side by side, no setup required.
This page lists all 95 visual question answering models in the Roboflow Playground catalog: 36 open-weight models you can self-host and 59 proprietary models accessed through provider APIs; 88 of them run live in the Roboflow Playground on your own images. On our visual reasoning benchmark (63 models tested), Gemini 4 Argon by Google currently scores highest. Compare licenses, parameters, API prices, and release dates side by side, or open any model page for full details.