ERNIE 4.5-VL Thinking: Paper Review

Non-member link.

You know that moment when you stare at a screenshot or a messy chart and think,
“I wish the model could just… actually look at this instead of hallucinating around it”?

That’s basically the itch ERNIE 4.5-VL-Thinking is trying to scratch.

This isn’t “oh cool, it can caption cat photos” territory.
This thing is built to investigate images zoom in, reason step-by-step, call tools if needed, and behave more like a junior analyst than a sticker-making camera filter.

In this post, let’s walk through:

• What ERNIE 4.5-VL-Thinking actually is (in human terms)
• Where it shines: concrete use cases you’d actually build
• The different “flavours” in the ERNIE 4.5-VL family and when to use which
• How it does multimodal reasoning under the hood

I’ll walk you like a person who’s also tired of hype and just wants to ship things.

So, what is ERNIE 4.5-VL-Thinking, really?

A 30B-param multimodal MoE model where only ~3B parameters are active per token, tuned heavily for reasoning with images, documents, charts and video, and released under Apache-2.0.