Moondream 3 Preview: Frontier-Level Reasoning at a Blazing Speed

That’s actually kinda impressive for an 8b model. Normally my experience with them is that they’re not really useful.

I tried it out on their website and it seems pretty legit, it gets stuff wrong but so do all the vision models in my experience

Moondream 2 has been very useful for me: I've been using it to automatically label object detection datasets for novel classes and distill an orders of magnitude smaller but similarly accurate CNN.

One oddity is that I haven't seen the claimed improvements beyond the 2025-01-09 tag - subsequent releases improve recall but degrade precision pretty significantly. It'd be amazing if object detection VLMs like this reported class confidences to better address this issue. That said, having a dedicated object detection API is very nice and absent from other models/wrappers AFAIK.

Looking forward to Moondream 3 post-inference optimizations. Congrats to the team. The founder Vik is a great follow on X if that's your thing.

Impressive stuff! Has anyone tried it for computer/browser control? How does it fare with graphs and charts?

Spent 5 minutes trying to get basic pricing info for Moondream cloud. Seems it simply does not exist (or at least not until you've actually signed up?). There's 5,000 free requests but I need to sense-check the pricing as viable as step 0 of evaluating - long before hooking it up to an app.

This looks amazing. I'm a big fan of Gemini for bounding box operations, the idea that a 9B model could outperform it is incredibly exciting!

I noticed that Moondream 2 was Apache 2 licensed but the 3 preview is currently BSL ("You can’t (without a deal): offer the model’s functionality to anyone outside your organization—e.g., an external API, or managed hosting for customers") - is that a permanent change to your licensing policies?

Really impressive performance from the Moondream model, but looking at the results from the big 3 labs, it's absolutely wild how poorly Claude and OpenAI perform. Gemini isn't as good as Moondream, but it's clearly the only one that's even half way decent at these vision tasks. I didn't realize how big a performance gap there was.

Can anyone suggest what's the cheapest hardware to run this model locally with a reasonable performance?

Would be interesting to see how it scores on COCO or Object356 dataset object detection (even if I know will be slower than dedicated object detection model)

Is there a GGUF?

It's ability to process large volumes of images with low active parameters makes it a significant advancement for edge devices. However, scaling these models to production environments often introduces security challenges, including bot floods targeting inference APIs and adversarial inputs that mimic legitimate queries to disrupt detections.

So... it should be really good at ARC?

Could you clarify whether the 2B active parameter concept refers to per-token inference and how this scales with context length? Specifically how MoE affects activation during inference and any practical implications for latency.

Any examples of face tracking a video and exporting mask coordinates to apply a filter in ffmpeg

Using moondream2 at paper.design to describe user uploaded images (for automatic labels in the layer tree). It's incredible, super fast and accurate. Excited to try out 3 :)

How does it perform against the new Qwen3-VL model?

The MoE architecture choice here is particularly interesting - the ability to keep only 2B parameters active while maintaining 8B model performance is a game-changer for edge deployment. I've been deploying vision models in production environments where latency is critical, and this sparse activation approach could solve the inference cost problem that's been limiting adoption of larger VLMs. The chart understanding capabilities mentioned look promising for automated document analysis workflows. Has anyone tested the model's consistency across different image qualities or lighting conditions? That's often where smaller models struggle compared to frontier ones.

it's honestly really good. Big fan of that team, they are really practical and have been producing really useful software and sharing all their learnings online.

Tried it's detection out on the playground as a 9B model it's pretty good.