Best Vision Language Models
Compare the best 69 vision language models and try 60 of them on your own image, free in the Roboflow Playground. 32 are open-weight, so you can self-host them for free under their licenses.
69 models · 32 open-weight · 60 free to try · prices synced Jul 24, 2026
Best Vision Language Models 2026
The best vision language model on our vision benchmark across six tasks right now is Gemini 3.5 Flash by Google, scoring 87.1% across 21 models tested, followed by Gemini 3.1 Pro at 86.3%. Updated Jul 24, 2026. See the full ranking.
Benchmark Leader
Gemini 3.5 Flash
Highest score on our vision benchmark across six tasks, 21 models tested
Open-Weight Leader
Qwen3-VL 235B
Best score among the open-weight models, #15 of 21 overall on our vision benchmark across six tasks
Fastest on Benchmark
Gemini 3.5 Flash-Lite
Quickest in our vision benchmark across six tasks runs: 2.7s per request on average
Lowest Measured Cost
Qwen3-VL 235B
Cheapest in our vision benchmark across six tasks runs: $0.0006 per request as measured, not list price
Benchmark scores
Ranked on our vision benchmark across six tasks, 21 models tested, updated Jul 24, 2026
Open-Source Vision Language Models
32 models with downloadable weights you can self-host under their licenses (Modified MIT, Apache 2.0, and MIT). 23 run live in the Playground through hosted APIs, so self-hosting is optional.
Kimi K3Moonshot AI | 2.8T | Modified MIT | Jul 2026 | |
Gemma 4 12BGoogle | 12B | Apache 2.0 | Jun 2026 | |
Qwen3.6 27BQwen | 27B | Apache 2.0 | Apr 2026 | |
Qwen3.6 35B A3BQwen | 35B total, 3B active | Apache 2.0 | Apr 2026 | |
Gemma 4 26B A4BGoogle | 25.2B | Apache 2.0 | Apr 2026 | |
Gemma 4 31BGoogle | 31B | Apache 2.0 | Apr 2026 | |
GLM-OCRZ.ai | 0.9B | MIT | Mar 2026 | |
Qwen3.5 9bQwen | 9B | Apache 2.0 | Mar 2026 |
Vision Language Models via API
37 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.
Gemini 3.5 Flash-LiteGoogle | $0.30 | $2.50 | 1.0M | Jul 2026 | |
Gemini 3.6 FlashGoogle | $1.50 | $7.50 | 1M | Jul 2026 | |
GPT-5.6 LunaOpenAI | $1.00 | $6.00 | 1.5M | Jul 2026 | |
GPT-5.6 SolOpenAI | $5.00 | $30.00 | 1.5M | Jul 2026 | |
GPT-5.6 TerraOpenAI | $2.50 | $15.00 | 1.1M | Jul 2026 | |
Muse Spark 1.1Meta | $1.25 | $4.25 | 1.0M | Jul 2026 | |
Claude Sonnet 5Anthropic | $2.00 | $10.00 | 1M | Jun 2026 | |
Claude Fable 5Anthropic | $10.00 | $50.00 | 1M | Jun 2026 |
How to Choose a Vision Language Model
Every model on this page can describe, read, and answer questions about images; the differences that matter in practice are capability ceiling, cost per call, latency, and whether you can self-host. Four questions settle most decisions.
Frontier models from the major labs hold a real edge on multi-step reasoning, dense documents, and unusual images, and that edge is measurable: our vision benchmark on the evals page scores the models we have tested across six task types. If your workload lives near the ceiling (complex documents, open-ended analysis), the frontier tier earns its price. If it is mostly captioning, tagging, and simple questions, mid-tier and small open models are dramatically cheaper for nearly the same output.
Open-weight VLMs (the Qwen VL line, Gemma, Llama vision models) can run on your own hardware: no per-call cost, no data leaving your infrastructure, no rate limits, at the price of operating GPUs yourself. Closed models via API invert that tradeoff. Privacy requirements, offline deployment, and predictable high volume favor open weights; spiky or low volume and access to the strongest models favor APIs.
A VLM is the flexible generalist, but production systems often beat it on cost and accuracy by decomposing: a trained detector localizes, an OCR model reads, and the VLM only reasons over the crops or the extracted text. Use the VLM to prototype the whole task, then replace the fixed, high-volume parts with specialized models as requirements settle.
The bottom line: Match the tier to your hardest task, choose open weights when privacy or unit economics demand it, and treat the VLM as the reasoning layer rather than the whole pipeline once volume grows.
Frequently Asked Questions About Vision Language Models
Vision language models (VLMs) are general-purpose models that join a vision encoder to a large language model, so one system can describe images, answer questions about them, read text in them, and reason across image and prompt together. The vision side encodes the image into tokens the language side can attend to, which is why the same model handles captioning, VQA, OCR-style reading, and detection-flavored questions zero-shot from instructions. GPT, Claude, Gemini, and Qwen VL are examples. VLMs trade some per-task accuracy, speed, and cost against enormous breadth, so the common production pattern is to prototype with a VLM and move fixed, high-volume tasks to a specialized or fine-tuned model once requirements settle. This page lists 69 vision language models, including 32 open-weight options you can self-host; 60 of them run live in the Playground so you can test them on your own images.
On our vision benchmark across six tasks (21 models tested, updated Jul 24, 2026), Gemini 3.5 Flash by Google currently scores highest at 87.1%. 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. 32 of the 69 vision language models here are open-weight (for example Kimi K3, Gemma 4 12B, and Qwen3.6 27B), free to self-host under their licenses (Modified MIT, Apache 2.0, and MIT).
Yes. You can run 60 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 69 vision language models in the Roboflow Playground catalog: 32 open-weight models you can self-host and 37 proprietary models accessed through provider APIs; 60 of them run live in the Roboflow Playground on your own images. On our vision benchmark across six tasks (21 models tested), Gemini 3.5 Flash by Google currently scores highest. Compare licenses, parameters, API prices, and release dates side by side, or open any model page for full details.