Updated
Vision LLMs read images, screenshots, charts, and PDFs, then answer in text. The hard part is matching one to your job: a cheap high-volume reader and a frontier document-reasoner sit far apart on price, speed, and accuracy. These 17 picks cover both ends of that range.
Best Vision LLMs
#ModelBest for
How to Choose
When choosing between these models, consider:- Access: Decide first whether you’ll use the model in an app, call it through an API, or run it locally, because that single choice drives cost, privacy, latency, and setup work more than small score differences do. Qwen3.5 4B runs on a typical laptop; Gemma 4 31B and Qwen3.6 27B need high-end local hardware. Open-weight leaders like Kimi K2.6 and MiniMax-M3 need self-hosting infrastructure, not a workstation.
- Quality: We use a vision score that blends Arena’s vision arena (human preference, style-controlled) with Artificial Analysis’s MMMU-Pro visual reasoning, normalized to a percentage. Two caveats matter. Claude Opus 4.8 and Grok 4.5 carry observed-only scores that aren’t directly comparable to the fully benchmarked models above them, and the newest premium models, Claude Fable 5 and GPT-5.6 Sol, rank lower than their reputations suggest mainly because standardized vision coverage lags their release.
- Price: We compare USD per 1,000 one-megapixel images at 1024x1024, image input only. It’s the cleanest way to line up costs, though your real bill also depends on the text tokens each request generates.
- Vision Latency: Time to first token for one image plus roughly 1,000 input tokens, where lower is better. It captures responsiveness, not throughput. Gemini 3.5 Flash is slow to first token but built for high-volume batches, so match the metric to how you’ll actually use the model.
Other Models We Considered
Frequently Asked Questions
What is the best vision LLM right now?
What is the best vision LLM right now?
For peak document and diagram accuracy, Claude Opus 4.7 leads. For the best mix of accuracy and price, Gemini 3.5 Flash is the default recommendation; choose GPT-5.5 when first-response latency matters more.
What is the best vision LLM for most people?
What is the best vision LLM for most people?
Gemini 3.5 Flash covers the widest range of everyday image and document work cheaply and well. If you want Anthropic’s careful reading at a moderate price, Claude Sonnet 5 is the close alternative.
What is the best cheap or free vision LLM?
What is the best cheap or free vision LLM?
Gemma 4 31B has no model-usage fee when run locally on a high-end machine and is currently free through a hosted route. Qwen3.5 4B costs almost nothing and runs on a normal laptop, while MiniMax-M3 is the cheapest capable paid hosted option.
What is the best vision LLM you can run locally?
What is the best vision LLM you can run locally?
Qwen3.5 4B is the only pick here that runs on a typical laptop. If you have a high-end GPU, Gemma 4 31B is the stronger local choice.
What is the best open-weight vision LLM?
What is the best open-weight vision LLM?
Kimi K2.6 is the strongest open-weight model on this list, though it’s large enough that most people will use it hosted rather than self-hosted.
Is Gemini 3.5 Flash better than GPT-5.5?
Is Gemini 3.5 Flash better than GPT-5.5?
It depends on the job. Flash is much cheaper and is the better-value batch option; GPT-5.5 returns a much faster first response. For interactive tools, GPT-5.5; for high-volume pipelines, Flash.
Do vision benchmarks match real-world use?
Do vision benchmarks match real-world use?
Roughly. They track document reading and visual reasoning well, but they don’t capture your exact images, latency needs, or task mix. Test the top two or three candidates on your own inputs before committing.
What should I use instead of GPT-4o?
What should I use instead of GPT-4o?
GPT-5.5 is the direct upgrade for fast, high-detail document reading. If cost and volume matter more, Gemini 3.5 Flash is the better move.