baidu/ERNIE-4.5-VL-28B-A3B-Thinking released. Curious case..

Posted by PaceZealousideal6091@reddit | LocalLLaMA | View on Reddit | 19 comments

It seems Baidu has released the "thinking" variant if their vl model silently. The earlier model was supposedly hybrid, supporting both "thinking" and "non-thinking". The model card says that they have introduced something called "thinking with images" without explaining what it is. They have one put a small hardly visible graph comparing it with gemini 2.5 pro and gpt-5 high in various benchmarks . If you squint your eye enough, then you'll see they claim using the graph that this model keeps up or beat them good in many of the benchmarks. Surely benchmaxxed. Its too good to believe. Has anyone tried it? The previous ernie versions have been decent. It might be worth testing it. Does anyone have any idea how is this "thinking" variant different?

Reply to Post

19 Comments

[-]

thebadslime@reddit

I use the 21b constantly, favorite local model.

[-]

PaceZealousideal6091@reddit (OP)

How is it? I am assuming 21B is text only model. What are ir use-cases and have you tried Qwen 3 30B A3B? How do they compare with each other?

[-]

thebadslime@reddit

Yes its text only, its neck and neck with qwen in benchmarks, but its faster and qwen gets stuck in thinking loops sometimes. I prefer no thinking models.

[-]

PaceZealousideal6091@reddit (OP)

I think you are confusing with the older qwen 2.5 30B A3B. The qwen 3 30B A3B has separate thinking and non-thinking variants. And yeah, I agree that the thinking variant was again than being useful. The new one is really fast and very good.

[-]

PhysicsAgreeable8170@reddit

I believe you are confusing with the older Qwen3 and Qwen2.5. There is not a "qwen 2.5 30B A3B". I believe you are saying the old hybrid Qwen3

[-]

PaceZealousideal6091@reddit (OP)

Yes! You are right! Things are moving so fast that its confusing 😅

[-]

ComputeVoid@reddit

Other interesting things from the model card: \> Through an extensive mid-training phase, the model absorbed a vast and highly diverse corpus of premium visual-language reasoning data. \> The model leverages cutting-edge multimodal reinforcement learning techniques on verifiable tasks \> Additionally, our innovative "Thinking with Images" feature, when paired with tools like image zooming and image search, dramatically elevates the model's ability to process fine-grained details and handle long-tail visual knowledge. All of this makes me wish there was deeper explanation! "Thinking with Images" sounds like they provided the model with image tools and did RL on verifiable tasks where the answer was "derivable" from the image. "Thinking" to me means test-time compute scaling: the model learns to iteratively use visual tools to produce variations on the input image, adding more vision tokens for harder problems.

[-]

PaceZealousideal6091@reddit (OP)

Has anyone got the vl variant gguf running on lcpp?

[-]

zenmagnets@reddit

Nope

[-]

Corporate_Drone31@reddit

I'm glad the ERNIE folks aren't tapping out, given all the other releases from other Chinese labs. We need diversity, and some of the ERNIE models I evaluated were quite promising.

[-]

bobby-chan@reddit

from the model card, ("**Thinking with Images** 🤔🔍: The model thinks like a human, capable of freely zooming in and out of images to grasp every detail and uncover all information.) I wonder if it's like a multistep analysis of an image, with the zooming (probably more cropping) of different parts, to limit distraction? In-context-RAG like chunking?

[-]

Porespellar@reddit

“Zoom and enhance”, we finally have it!

[-]

Kathane37@reddit

o3 was using python to edit the image and zoom in on part that interested it. So i guess it will be similar and more an agentic fearure than some native skills.

[-]

CoruNethronX@reddit

https://preview.redd.it/ks6afm5uzl0g1.jpeg?width=460&format=pjpg&auto=webp&s=246c841578edf824e1a4cf45ca070ebad9243993

[-]

PaceZealousideal6091@reddit (OP)

Yeah... thats plausible. However, I wonder if it will shoot up the inference time for each images depending upon complexity. Waiting for a gguf to start testing it.

[-]

SlowFail2433@reddit

Ernie models are good especially their giant one that is over 400B. They used GSPO and IcePop. It is plausible they truly got a SOTA for their size here.

[-]

Mir4can@reddit

Are there any benchmark scores released?

[-]

PaceZealousideal6091@reddit (OP)

Yeah. If you call it one. You can see it in the model card from the HF link I shared. Its so small that its difficult to understand it properly.

[-]

Mir4can@reddit

Oh just saw that. Thanks