[HN Gopher] Moondream 3 Preview: Frontier-level reasoning at a b...
___________________________________________________________________
Moondream 3 Preview: Frontier-level reasoning at a blazing speed
Author : kristianp
Score : 271 points
Date : 2025-09-26 21:59 UTC (1 days ago)
(HTM) web link (moondream.ai)
(TXT) w3m dump (moondream.ai)
| Aeolun wrote:
| That's actually kinda impressive for an 8b model. Normally my
| experience with them is that they're not really useful.
| conwayanderson wrote:
| Only 2b active also - very fast
| lawlessone wrote:
| Can run it on a phone then?
|
| Seems like it could be somewhat useful for people with poor
| eyesight or blindness
| conwayanderson wrote:
| In terms of size yes, but I think it needs some work to get
| the model in the right format
|
| couple people got it running on a raspberry pi though
| apwell23 wrote:
| sorry what does it mean for only 2b to be active?
| simonw wrote:
| My understanding is that, while all 8B are loaded into
| memory, for each token inference step only 2B are selected
| and used - so tokens are produced faster because there is
| less computation needed.
|
| Hoping someone will correct me if that's not the right
| mental model!
| derac wrote:
| 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
| conwayanderson wrote:
| Especially good at detection & pointing cases - especially
| since the bigger models aren't good at localization
| scoots_k wrote:
| 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.
| conwayanderson wrote:
| Also used it for auto-labeling - it's crazy good for that
| radq wrote:
| Thanks! If you could shoot me a note at vik@m87.ai with any
| examples of the precision/recall issues you saw I'd appreciate
| it a ton.
| scoots_k wrote:
| Will do!
| nstj wrote:
| Wonderful to see "at the coalface" collaboration happen on
| this stuff at HN. More than just a newsfeed!
| buyucu wrote:
| are you planning to release a GGUF?
| sheepscreek wrote:
| Impressive stuff! Has anyone tried it for computer/browser
| control? How does it fare with graphs and charts?
| radq wrote:
| The 'point' skill is trained on a ton of UI data; we've heard
| of a lot of people using it in combination with a bigger driver
| model for UI automation. We are also planning on post-training
| it to work end-to-end for this in an agentic setting before the
| final release -- this was one of the main reasons we increased
| the model's context length.
|
| Re: chart understanding, there are a lot of different types of
| charts out there but it does fairly well! We posted benchmarks
| for ChartQA in the blog but it's on par with GPT5* and slightly
| better than Gemini 2.5 Flash.
|
| * To be fair to GPT5, it's going to work well on many more
| types of charts/graphs than Moondream. To be fair to Moondream,
| GPT5 isn't really well suited to deploy in a lot of vision AI
| applications due to cost/latency.
| bobdyl87 wrote:
| Im labeling a dataset with it. We'll see how it turns out
| bobdyl87 wrote:
| Pretty good so far. Have 100,000 detections
| stephenbuilds wrote:
| 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 :)
| robertdaniels wrote:
| 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.
| Onavo wrote:
| How does it perform against the new Qwen3-VL model?
| kache_ wrote:
| 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.
| buyucu wrote:
| Is there a GGUF?
| liqilin1567 wrote:
| Tried it's detection out on the playground as a 9B model it's
| pretty good.
| Imanari wrote:
| So... it should be really good at ARC?
| pzo wrote:
| 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)
| bluelightning2k wrote:
| 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.
| civilchaos wrote:
| We are looking to launch our cloud very soon. We are still
| optimizing our inference to get you the best pricing we can
| offer. Follow @moondreamai on X if you want your ear to the
| ground for our launch!
| aitchnyu wrote:
| Will you add this to OpenRouter too?
| nicohayes wrote:
| 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.
| simonw wrote:
| 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?
| simonw wrote:
| I just noticed in
| https://huggingface.co/moondream/moondream3-preview/blob/mai...
| that the license is set to change to Apache 2 after two years.
| nicohayes wrote:
| 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.
| ZeroCool2u wrote:
| 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.
| ekidd wrote:
| Gemini is really fantastic at anything that's OCR-adjacent, and
| it promptly falls over on most other image-related tasks.
| Jackson__ wrote:
| Funnily enough, Gemini is also the only one able to read a D20.
| ChatGPT consistently gets it wrong, and Claude mostly argues it
| can't read the face of the die that's facing up because it's
| obstructed (it's not lol).
| KronisLV wrote:
| I'm not sure why they haven't been acquired yet by any of the
| big ones, since clearly Moondream is pretty good! Definitely
| seems like something Anthropic/OpenAI/whoever would want to
| fold into their platforms and such. Everyone involved in
| creating it should probably be swimming in money and visual use
| cases for LLMs should become far less useless with the reach of
| the big orgs.
| thw_9a83c wrote:
| Can anyone suggest what's the cheapest hardware to run this model
| locally with a reasonable performance?
| daemonologist wrote:
| Since there's no quantized version available at the moment,
| you'll need ~20 GB of memory for the weights plus some extra
| for the KV cache. CPU with 32 GB RAM will be the cheapest and
| still reasonably fast given the relatively small number of
| activated parameters.
| thw_9a83c wrote:
| Thank you!
|
| I don't even know what a "quantized version" is, but I was
| expecting answers about NVIDIA graphics cards and their
| memory. My computer has 24GB of memory, but I'll go for 64GB
| to run this locally on a new computer.
___________________________________________________________________
(page generated 2025-09-27 23:01 UTC)