[HN Gopher] FastVLM: Efficient vision encoding for vision langua...
       ___________________________________________________________________
        
       FastVLM: Efficient vision encoding for vision language models
        
       Author : nhod
       Score  : 349 points
       Date   : 2025-05-13 01:16 UTC (21 hours ago)
        
 (HTM) web link (github.com)
 (TXT) w3m dump (github.com)
        
       | BryanLegend wrote:
       | Seems like the main thing holding these new minds back is being
       | able to see well. Breakthroughs like this will fix that.
        
         | efnx wrote:
         | That and the ability to hold on to knowledge.
        
           | static_void wrote:
           | ... or say they don't know.
        
       | kamranjon wrote:
       | Apple out here playing 5d chess, installing neural cores in their
       | hardware and writing crazy efficient vision models to run on em.
       | Cool stuff.
        
         | wmf wrote:
         | I thought they turned sycophancy off...
        
           | kamranjon wrote:
           | Awe yes I admit, I think the new Apple hardware is real cool
        
       | vFunct wrote:
       | Can it fill a wine glass to the rim?
        
         | mkl wrote:
         | It's for interpreting images, not generating them.
        
       | turnsout wrote:
       | Apple has gotten a slow start in the LLM world, but they have the
       | only long term strategy that makes sense. They're going to
       | dominate the 2030s.
        
         | boroboro4 wrote:
         | What exactly the strategy is?
        
           | generalizations wrote:
           | They can run locally on-device: a win for cost, latency and
           | privacy (privacy is pragmatic: it means you can use all the
           | user's data as context without qualms). There's a reason
           | Microsoft tried so hard to push for the neural processors a
           | year or two ago. Avoiding the cost of the datacenter while
           | offering good-enough inference (emphasis on good) is a
           | massive win.
        
             | turnsout wrote:
             | Yes, thank you; this is the strategy I was referring to. It
             | will take some time for the models and chips to get there,
             | but on-device inference will have massive advantages for
             | privacy, speed and cost. Plus it will drive demand for
             | hardware--at first, iPhones, but soon AirPods and glasses.
        
             | xnx wrote:
             | Google already has some of the best on device models
             | (Gemma) and chips (Tensor).
        
               | AceJohnny2 wrote:
               | > _and chips (Tensor)_
               | 
               | Is there actually any hard data out there comparing the
               | NPU on the Google Tensor G4 vs the Apple A18? I wasn't
               | able to quickly find anything concrete.
               | 
               | I mean Apple has been shipping mobile NPUs for longer
               | than Google (Apple: since A11 in 2017, Google: since
               | 2021), and are built on (ostensibly) a smaller silicon
               | node that Google's (G4: Samsung SF4P vs A18: TSMC N3E).
               | However, the G4 appears to have more RAM bandwidth (68.26
               | GB/s vs 60 GB/s on A18).
        
               | lern_too_spel wrote:
               | Google has been shipping custom NPUs since the Pixel 4 in
               | 2019. Prior to that, Google phones just used off the
               | shelf SOCs from Qualcomm, with 2018's Pixel 3 using the
               | NPU in the Snapdragon 845. Android first shipped NNAPI in
               | Android 8.1 in 2017, with acceleration on various mobile
               | GPUs and DSPs, including the Pixel Visual Core on the
               | Pixel 2. Google has shipped more on-device models so far,
               | but neither company has a moat for on-device inference.
               | 
               | https://blog.google/products/pixel/pixel-visual-core-
               | image-p...
        
               | turnsout wrote:
               | Unfortunately for them, Google doesn't make devices that
               | people want to buy
        
               | lern_too_spel wrote:
               | Other Android phone vendors do, and they have the same
               | strategy, sitting on top of Qualcomm NPUs.
        
             | weikju wrote:
             | They are running data centers and offloading some things to
             | chatGPT though, not just running on device.
             | 
             | In fact there's no clear indication when Apple Intelligence
             | is running on-device or in their Private Cloud Compute.
        
         | jfarina wrote:
         | What strategy is that?
        
           | ryanmcgarvey wrote:
           | I presume they mean that distribution is king and they make
           | all the devices.
        
         | karn97 wrote:
         | Average delusional hner who really doesn't know what they are
         | talking about
        
           | turnsout wrote:
           | Enlighten us, wise one
        
       | insane_dreamer wrote:
       | As the father of a young child whose optic nerves are highly
       | deteriorated (compression) and is expected to lose his sight
       | (when exactly is unknown; based on original projections he should
       | be blind by now, but an experimental treatment run in a trial at
       | the NIH (KEEP FUNDING SCIENCE) has stabilized his sight), I'm
       | overjoyed with the advances being made in VLMs. I can now
       | envision a future where even if he loses his sight he'll be able
       | to interact with the world around him, go to college, have a
       | fulfilling career (he loves science and engineering, and is
       | talented for his young age), etc.
        
         | lynx97 wrote:
         | I grew up in the 80s as a 100% blind child. Technology was by
         | far not as advanced as today. Computers were just coming up
         | when I was around 12. I learnt to type on a oldschool
         | typewriter, and I also learnt to write braille with a pretty
         | heavy full-metal embossing device. OCR was still quite bad.
         | When I switched to what you call high scooll, I used a laptop
         | with integrated Braille display to follow classes. Used good
         | old DOS as OS and Word 5.5 as my "notepad". Except for PC
         | Lingua for Latin, I basically had no tools specialized for
         | learning. A electronic notepad and my brain was all I had to
         | follow school. And I still made it. I have a great job I love,
         | my own appartment, a sweet girlfriend and I am basically
         | completely independent. To a point where I had to forcefully
         | send away my mother since her continued attempts to "help" me
         | were basically detrimental to my own development. I can not
         | emphasis how important it is how you deal with it as a parent.
         | Since parents are indeed the biggest hinderence to development,
         | we have a saying around here amongst disabled people:
         | "additional disability due to parental overprotection"
         | (Zusatzbehinderung Eltern). Please take a moment to understand
         | what this means, without feeling personally attacked. Its
         | important. Your child _can_ leave home around 18, just like
         | every other kid. I did. Don 't slow that process down
         | artificially. The more this is prolonged, the harder it gets
         | for the individual to actually obtain independence.
         | 
         | I am telling you this because I read between the lines that you
         | believe current technology is a reason for you to be hopeful.
         | Sure, it should be. But never forget, your child can do much
         | more then you as a sighted person will ever be able to
         | understand. Don't let them drown in your own misery. Let them
         | discover what they can do. You will be surprised what they come
         | up with. And dont fall for Gear Acquision Syndrome. Sure, tools
         | are nice, and they do get better, which is also nice. I _LOVE_
         | vision models, to stay on topic somehow. However, I still leave
         | my house with only a cane and my phone in my pocket. I do
         | occasionally ask Siri  "Where am I" to get an address if I
         | happen to have forgotten where I am exactly, currently. But at
         | the end of the day, my cane is what shows me the way. Most tech
         | is hype, plain old hearing and your sense of touch gets you
         | much farther then you might think.
         | 
         | Wish you all the best for your own journey, and the development
         | of your child.
        
           | wiz21c wrote:
           | I should read a comment like yours every morning.
        
           | topato wrote:
           | Wow, this really adds an amazing perspective to the entire
           | (frequently touted) concept of Visual Language Models somehow
           | "saving" blind people from their old life; In the past, a
           | blind person desperately needed caretakers, otherwise the
           | blind person will bumble around their home, end up mistaking
           | the sink for the toilet, accidentally turn on their stove
           | thinking it's the thermostat, until they died after mistaking
           | bleach for milk and cat litter for cereal....
           | 
           | BUT NOW... THE FUTURE IS HERE.... an all-knowing god-like
           | cell phone can tell these poor miserable individuals what the
           | objects in their own homes are! No more tragic Mr. Magoo-ian
           | accidents!
           | 
           | But thank you for posting this; It certainly enlightened me!
           | I'll admit, all these AI solutions
        
             | exe34 wrote:
             | > I'll admit, all these AI solutions
             | 
             | They got to him.
        
           | insane_dreamer wrote:
           | Thanks. I appreciate your insight.
        
       | liamwire wrote:
       | It feels like this is the required level of speed-up needed re.
       | time-to-first-token to make continuous vision useful for on-
       | device applications like an assistant that can see and take
       | action on your screen, ala the original Apple Intelligence demos.
       | It's very impressive seeing the app in the repo and I'm excited
       | to build it tonight and play around.
        
       | nine_k wrote:
       | With that, a really helpful aid for blind people can be made,
       | running just on their phone, fed from a camera in their
       | eyeglasses. Somebody who could not move around without an
       | assistant could become autonomous in daily life.
        
         | jdiff wrote:
         | It might be useful for telling Cream of Chicken from Cream of
         | Mushroom, but for locomotion I can't see this adding anything
         | over existing strategies people use to get around sans sight.
         | 
         | "There's a tree. There's a tree. There's a tree. There's a
         | number of pedestrians. There's a tree. There's a sign." does
         | not strike me as useful feedback for getting around.
        
           | nine_k wrote:
           | Consider a city. It's full of signs and inscriptions, traffic
           | lights, and other key interaction elements. Consider a store.
           | It has shelves with stuff, again with inscriptions, price
           | tags, etc.
           | 
           | "Pavement. Row of stores to the left. Joe's Grocery Store.
           | Doors. Door handle. A shelf with bakery. A shelf with canned
           | goods. A shelf with bottles. Coke bottle. Large Pepsi bottle.
           | Apple juice bottle. Passageway. Checkout. Payment terminal.
           | Door. Door handle. Pavement. ..."
        
             | jdiff wrote:
             | None of that gives me any useful spatial sense of _where_.
             | "Payment terminal." Okay. Where is it? Left? Left where?
             | How much left? How far?
             | 
             | The only truly useful bits I see in your stream of text is,
             | again, "Cream of Mushroom" vs "Cream of Chicken." I am
             | actively holding something, so I know where it is, but need
             | to differentiate it from printed detail.
        
       | adamsiem wrote:
       | Anyone using vision to parse screenshots? QVQ was too slow. Will
       | give this a shot.
        
         | abrichr wrote:
         | You might be interested in
         | https://github.com/OpenAdaptAI/OpenAdapt
        
         | logankeenan wrote:
         | I used molmo to parse screenshots in order to detect locations
         | of UI elements. See the repo below. I think Omni parser from
         | Microsoft would also work well.
         | 
         | https://github.com/logankeenan/george
         | 
         | https://github.com/microsoft/OmniParser
        
       | Aeroi wrote:
       | I built/building a realtime voice+vision app called Sen, its
       | currently live in beta and streams frames over webrtc. It's fast
       | and smart, but Im super curious to see how these models do as we
       | get closer to the metal. I can see these running on-device in the
       | future with super fast ttfb.
        
         | keyle wrote:
         | Do you have a write up of the tech stack and setup? Or willing
         | to give the gist here?
         | 
         | I'd like to make a private Qwen or similar for my kids to
         | prompt with a button and voice control. It doesn't need
         | vision... Although eventually that'd be very cool.
         | 
         | Siri just sucks.
         | 
         | We might not be there yet...
        
           | Aeroi wrote:
           | yeah i made a post on here, but the algo sent it to the gulag
           | abyss.
           | 
           | https://news.ycombinator.com/item?id=43926673
        
             | keyle wrote:
             | That's a good product site but it doesn't help me in
             | anyway...
        
           | Aeroi wrote:
           | I also ran across an interesting robot toy demo today that
           | had voice built in. it was whimsical and seemed like it was
           | aimed towards primary education and kids. Someone here might
           | know the name.
        
           | stavros wrote:
           | You can use Ollama or LM Studio, both in API mode, to return
           | the responses. I believe they offer audio support, but I'm
           | not entirely sure.
           | 
           | However, if you're looking for instruction following (like an
           | agent), I've tried to implement my own agent and have lost
           | faith. Even GPT-4.1 will regularly gaslight me that no, it
           | _definitely_ ran the tool call to add the event to my
           | calendar, when it just didn 't. I can't get any more
           | adherence out of it.
        
           | cloudking wrote:
           | Check out https://livekit.io/
        
           | tomp wrote:
           | We're definitely there, there's just no "ready-made" apps
           | yet. But the technology is possible, go to e.g. vapi.ai to
           | test it.
        
       | nikolayasdf123 wrote:
       | 2GB for 0.5B smallest model. it does not make sense for each app
       | to download this. apple must have plans to pre-load these models
       | on os level and expose SDK for all apps to call these models
       | locally. exciting times!
       | 
       | opened issue for them to confirm this:
       | https://github.com/apple/ml-fastvlm/issues/7
        
         | cube2222 wrote:
         | That's what they suggested about LLMs at last year's WWDC iirc.
         | The core models are provided by the OS, while apps bring LORAs
         | to fine-tune them / bring custom heads for them.
        
         | babl-yc wrote:
         | You could probably get away with f16 or even quantize to int8
         | and have a much smaller model, but your point stands. Users
         | won't be thrilled to download a 500MB model for an app either.
        
           | nikolayasdf123 wrote:
           | haha latest Uber build for iOS 18 is 500MB... without LLM
           | models <face-palm/>
        
             | ukuina wrote:
             | What are they doing in there? Is it mostly visual assets?
        
               | bastawhiz wrote:
               | If I was going to guess, I'd get there's a ton of third
               | party code for things like payment method SDKs. Every
               | local payment method around the world is going to have
               | its own package that you need to import, and you can't
               | just load in new executable code on the fly after the app
               | is installed.
        
               | victorbjorklund wrote:
               | You can actually do over the air updates of apps (how
               | easy it is depends on what you wrote your app in) and not
               | adding a new feature (like just adding an additional
               | payment provider) would not require an update on the App
               | Store.
        
               | nikolayasdf123 wrote:
               | wouldn't you want to create payment gateway and abstract
               | away logic such that client is agnostic of payment
               | processes and backend confirms internal payment process
               | into external specific ones? (in worst case redirect to
               | other apps with universal links or webview)
        
               | philipkglass wrote:
               | A lot of geography-specific scenarios are compiled into
               | the app, including regional payment SDKs. There's a great
               | comment from a former Uber engineer explaining it here:
               | 
               | https://news.ycombinator.com/item?id=25376346
        
               | nikolayasdf123 wrote:
               | I think they using vector graphics and vector animations
               | (say Rive). Rive takes order of 10s of KBs. Lottie is
               | much larger to 100s of KBs. Even then you would need 5000
               | animations to reach 500MB, unlikely!
               | 
               | raster graphics and videos are likely not included in
               | build
               | 
               | probably some unused code (libraries) got into it, it can
               | grow quite large
               | 
               | or maybe some ML models?
        
         | HanClinto wrote:
         | I think that there is fantastic potential in having open-
         | weight, OS-standard foundation models.
         | 
         | Especially if the API gives opportunity for app developers to
         | load their custom LoRa fine-tunings onto OS-standard foundation
         | models at runtime, then you can (ideally) have the best of both
         | worlds -- fine-tuned app-specific models with reasonable app
         | sizes.
        
           | HappMacDonald wrote:
           | I haven't seen much done with loras for LLMs though, only for
           | diffusion image gen models. From what I've heard it sounds
           | like a difference in benefit due to architecture.
        
         | gessha wrote:
         | My guess is that they won't confirm it unless it's a big
         | presentation. WWDC maybe?
        
       | nikolayasdf123 wrote:
       | google and cloud LLM providers must be biting their teeth now!
       | haha
        
       | nikolayasdf123 wrote:
       | distributing this heavy compute and moving it close to device
       | where 1. source of data happens; 2. decision and output about the
       | result of analysis is done; is way to go. super low latency, no
       | network traffic, privacy, less overhead in cloud. this is amazing
        
       | porphyra wrote:
       | It seems that the future of robotics is VLA models. Even Tesla
       | FSD is an end-to-end VLA model. Efficient vision encoding will be
       | a huge part of making robots safe and responsive.
        
       | lynx97 wrote:
       | I wonder, can I convert/run this with llama.cpp? It being LLaVA
       | based seems promising.
        
       | vessenes wrote:
       | Um wow. The on-device realtime videos are worth a watch, and
       | compelling. Looking forward to this being deployed and widely
       | adopted. Getting much faster time to first token opens up a ton
       | of features and usability benefits.
        
       | buyucu wrote:
       | where is my gguf?
        
       | simianparrot wrote:
       | I have a feeling feeding tesseract the image every 1 second would
       | be significantly faster and take far less space and processing
       | power? Haven't tested it yet but given how fast tesseract is on
       | large images, it wouldn't surprise me.
        
         | regularfry wrote:
         | If all you want is OCR, possibly.
        
           | coredog64 wrote:
           | If all you want is OCR of typewritten text.
           | 
           | Tesseract is awful for handwriting.
        
       | d3k wrote:
       | Very nice! I wish they were more keen to contribute to AI/ML
       | community an publish weights and model definition on HuggingFace.
       | Funny enough I have just seen today a similar demo that is using
       | a freely available VLM: https://github.com/ngxson/smolvlm-
       | realtime-webcam
        
         | tough wrote:
         | SmolVLM is from huggingface team
         | 
         | cool to see people doing stuff with smaller models
         | 
         | https://huggingface.co/blog/smolvlm
         | 
         | https://arxiv.org/abs/2504.05299
        
       | labadal wrote:
       | I'm absolutely thrilled that there is an effort to make models
       | smaller and run with less resources instead of blindly throwing
       | more resources at the problem and expecting it to get solved.
        
       ___________________________________________________________________
       (page generated 2025-05-13 23:02 UTC)