Roboflow

Best LLMs with Vision Capabilities

Compare the best 97 vision LLMs and try 86 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.

97 models · 35 open-weight · 86 free to try · prices synced Nov 18, 2026

Best Vision LLMs 2026

GPT-6.2 Sol scores 90.3% on Roboflow's Vision Evals across five tasks (72 models), ahead of GPT-6.2 Astra at 90.2%. Updated Nov 17, 2026. See the full ranking.

Open-Source Vision LLMs

35 models with downloadable weights you can self-host under their licenses (Custom, MIT, and Apache 2.0). 27 run live in the Playground through hosted APIs, so self-hosting is optional.

Actions
816B total / 42B activeCustomOct 2026
Try
Mistral
1.05T total, 49B activeCustomOct 2026
Try
309B total, 15B activeMITSep 2026
Try
1.02T total, 42B activeMITSep 2026
Try
320B total, 18B activeMITAug 2026
Try
Qwen
27.78BApache 2.0Aug 2026
Try
29.6BApache 2.0Aug 2026
Try
MoonshotAI
Kimi K3Moonshot AI
2.8TModified MITJul 2026
Try

Vision LLMs via API

62 proprietary models where the weights aren't downloadable: access is through each provider's API and billed by them. Try 59 of them free in the Playground.

Actions
GoogleNEW
———Nov 2026
Try
GoogleNEW
———Nov 2026
Try
AnthropicNEW
$8.00$40.001MNov 2026
Try
MetaNEW
$0.10$0.501.0MNov 2026
OpenAI
$1.00$5.001.1MOct 2026
Try
Google
$2.00$10.001MOct 2026
Try
Meta
$4.00$20.001.0MOct 2026
Try
Anthropic
$0.10$0.501MOct 2026
Try

When Do You Need a Vision LLM?

The expensive mistake in either direction: using a frontier vision LLM for a job a small trained model does better and a thousand times cheaper, or spending weeks building a training pipeline for a job a VLM handles with one prompt. The decision comes down to how open-ended the task is.

If the task is fixed and repetitive (detect these products, classify these defects, count people in this camera feed), a trained detector or classifier wins on every production axis: per-image cost measured in fractions of a cent, millisecond latency, deployment on edge hardware, and consistency you can put error bars on. Vision LLMs answer the same question in seconds for orders of magnitude more money, with answers that can vary between calls.

Open-ended instructions, language output, and world knowledge are the VLM-only zone: describe this scene for an incident report, extract these fields from any invoice layout, answer whatever question a user types about an image. No trained classifier can do these, because the label space is unbounded. If the question changes per request, or the output is prose or structured text, you need the language model.

Within the family, capability tracks reasoning depth more than vision. Reading a receipt or writing alt text works on small open models at a fraction of frontier pricing; multi-step document reasoning and unusual images separate the top models quickly. Our vision benchmark on the evals page scores the models we have tested across five task types; the verdict module above names the current leader.

Mature production systems use the VLM as the reasoning layer, not the whole pipeline: a trained detector localizes at volume, OCR reads at volume, and the VLM handles the open-ended residue (judgment calls, summaries, escalations). Prototype everything with the VLM first, then move the fixed, high-volume steps to specialized models as the requirements stop changing.

The bottom line: Fixed question at volume: train a classic model. Open-ended question or language output: vision LLM. At scale, the answer is usually a pipeline with both.

Frequently Asked Questions About LLMs with Vision Capabilities

On Roboflow's Vision Evals across five tasks (72 models tested, updated Nov 17, 2026), GPT-6.2 Sol by OpenAI currently scores highest at 90.3%. 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 97 models here are open-weight (for example Qwen3.8 27B, Muse Spark Open, and Qwen3.6 27B), free to self-host under their licenses (Custom, MIT, and Apache 2.0).

Yes. You can run 86 of them in the Roboflow Playground for free. Upload an image and compare the models' output side by side, no setup required.

Explore other vision task pages, compare models head to head, or browse the full model catalog.