Best Vision Language Models
Compare the best 84 vision language models and try 75 of them on your own images, free in the Roboflow Playground. 35 are open-weight, so you can self-host them for free under their licenses.
84 models · 35 open-weight · 75 free to try · prices synced Sep 7, 2026
Best Vision Language Models 2026
The best vision language model on our vision benchmark across six tasks right now is GPT-6 Astra by OpenAI, scoring 86.6% across 53 vision language models tested, followed by Gemini 3.5 Flash at 86.0%. Updated Sep 5, 2026. See the full ranking.
Benchmark Leader
GPT-6 Astra
Highest score on our vision benchmark across six tasks, 53 vision language models tested
Open-Weight Leader
Qwen3.8 27B
Best score among the open-weight models, #17 of 53 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
Qwen3.7 Flash
Cheapest in our vision benchmark across six tasks runs: $0.0001 per request as measured, not list price
Benchmark scores
Ranked on our vision benchmark across six tasks, 53 vision language models tested, updated Sep 5, 2026
Open-Source Vision Language Models
35 models with downloadable weights you can self-host under their licenses (MIT, Apache 2.0, and Modified MIT). 26 run live in the Playground through hosted APIs, so self-hosting is optional.
| Actions | ||||
|---|---|---|---|---|
GLM 5.3 FlashZ.ai NEW | 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 | |
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 |
Vision Language Models via API
49 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 | |||||
|---|---|---|---|---|---|
GPT-6 AstraOpenAI NEW | $10.00 | $50.00 | 1.1M | Sep 2026 | |
Gemini 3.8 FlashGoogle NEW | $0.75 | $3.75 | 1.0M | Sep 2026 | |
Muse Spark 1.3Meta NEW | $1.25 | $4.25 | 1.0M | Sep 2026 | |
Claude Fable 5.1Anthropic NEW | $10.00 | $50.00 | 1M | Sep 2026 | |
Qwen3.8 FlashQwen NEW | $0.15 | $0.47 | 1M | Aug 2026 | |
Gemini 3.7 FlashGoogle | $0.75 | $3.75 | 1.0M | Aug 2026 | |
Grok 4.6SpaceXAI | $2.00 | $6.00 | 500K | Aug 2026 | |
Muse Spark 1.2Meta | $1.25 | $4.25 | 1.0M | Aug 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 84 vision language models, including 35 open-weight options you can self-host; 75 of them run live in the Playground so you can test them on your own images.
On our vision benchmark across six tasks (53 models tested, updated Sep 5, 2026), GPT-6 Astra by OpenAI currently scores highest at 86.6%. 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. 35 of the 84 vision language models here are open-weight (for example GLM 5.3 Flash, Qwen3.8 27B, and Muse Glimmer 30B), free to self-host under their licenses (MIT, Apache 2.0, and Modified MIT).
Yes. You can run 75 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 84 vision language models in the Roboflow Playground catalog: 35 open-weight models you can self-host and 49 proprietary models accessed through provider APIs; 75 of them run live in the Roboflow Playground on your own images. On our vision benchmark across six tasks (53 models tested), GPT-6 Astra by OpenAI currently scores highest. Compare licenses, parameters, API prices, and release dates side by side, or open any model page for full details.