Skip navigation
d1-2B by Liquid AI :Same model, three engines on a MacBook Pro M5 Max
0:00 / 0:48

d1-2B by Liquid AI :Same model, three engines on a MacBook Pro M5 Max

Ivan Fioravanti

@ivanfioravanti45K followers

Visit on X
305 likesDownloadOriginal post

17,801 views

d1-2B by Liquid AI :Same model, three engines on a MacBook Pro M5 Max - MLX: 66 ms, 153 W - Core AI GPU: 67 ms, 120 W - Core AI Neural Engine: 197 ms, 37 W The Neural Engine is 3x slower (on M5 Max, I bet this is different on M4) and draws a quarter of the power. Same answers on all three. I had to use a video instead of live camera feed because ANE usage goes up as soon as you start the camera. Core AI export from the original weights, plus live GPU / Neural Engine / power charts: Download and have fun! Share your stats on other Apple Silicon machines! I'm now creating an iPhone version 😎

37 Comments

Comments from the original post on X (Twitter)

  • Ivan Fioravanti's profile picture

    @ivanfioravanti

    @starkov100 has the solution for the ANE usage from camera!!!

    4 likes

  • Ivan Fioravanti's profile picture

    @ivanfioravanti

    @anemll can you try this on your Mac Mini M6 🙏 I'm curious to see results from new ANE!

    4 likes

  • Ivan Fioravanti's profile picture

    @ivanfioravanti

    New release data are slightly different 🤷🏻‍♂️ ⚡ MLX: 88 ms, 65 W ⚡ Core AI GPU: 78 ms, 68 W 🔋 Core AI Neural Engine: 183 ms, 37 W

    3 likes

  • Christopher Owen's profile picture

    @christopherowen

    @liquidai super cool!

    2 likes

  • Ivan Fioravanti's profile picture

    @ivanfioravanti

    @liquidai I'm creating the live deck of Local AI wonders! After this demo there is LFM3 to showcase differences!

  • Christopher Owen's profile picture

    @christopherowen

    @liquidai I cannot wait to see it! the 7 wonders of the modern world!

    1 like

  • Luis Catacora's profile picture

    @lucataco

    @liquidai d1-2B isnt the same as d1.2 right?

    2 likes

  • Ivan Fioravanti's profile picture

    @ivanfioravanti

    @liquidai It’s the previous, previous version.

  • Scott Miller's profile picture

    @scottmillerio

    @liquidai The camera bumping ANE usage is the detail I didn't expect. If the Neural Engine is shared with system features, a background model on it competes with the OS for time. Curious how steady that 197 ms stays with FaceTime or Continuity Camera running.

    1 like

  • Ivan Fioravanti's profile picture

    @ivanfioravanti

    @liquidai We have solution!

  • KC's profile picture

    @kcwolfy_

    @liquidai Thank you for being the first person to actually focus on the speed of a decision, one open source model versus just the result. Because between one second and three seconds for a great result misses the point. 👏 @liquidai and @ivanfioravanti

    1 like

  • Ivan Fioravanti's profile picture

    @ivanfioravanti

    @liquidai Thanks KC and thanks Liquid AI! Amazing model!

  • Yoshikawa Hitaka | 吉川飛空's profile picture

    @hitakayoshikawa

    Great comparison. For always-on workloads like voice dictation, that 37 W matters more than the extra 130 ms. I'm building Alice, a local Japanese dictation app on Apple silicon — curious whether the ANE gap shrinks for small STT encoders. Planning to test it on my M3 Ultra 512GB.

    1 like

  • Mladen Lotar's profile picture

    @lotar_vk

    @liquidai Wait... Camera uses ANE 👀? I mean awesome model and test, but camera confused me extremely 😅 Also ANE might be slower but if this is fast enough it uses 1/3 of energy which is awesome (and not slow at all)!

    1 like

  • Ivan Fioravanti's profile picture

    @ivanfioravanti

    @liquidai Not in this case, because I loaded a prerecorded video to avoid the problem. If you use live camera yes, it's using ANE and you'll get worst performance.

    1 like

  • Mladen Lotar's profile picture

    @lotar_vk

    Yeah, I figured it's a recording. Really got me confused. Now I somehow think Apple is silently tracking everything camera sees (what would be any sensible reason for it to fire up - guessing not video processing with all of the CPUs and GPUs). Tin foil hat on, time to figure it out 😂

  • @alexdolbun's profile picture

    @alexdolbun

    @liquidai Interesting direction, still not this 3ms decision like guys below 👇 done, but super impressive Mr. @ivanfioravanti GJ !! Reposted.

    1 like

  • Ivan Fioravanti's profile picture

    @ivanfioravanti

    @liquidai But they did it with text not video 😎 This is why I like d1!

    1 like

  • Norbert Klockiewicz's profile picture

    @Nklockiewicz

    @liquidai Can't wait for the iPhone numbers, on a phone a quarter of the power is probably worth more than the 3x 👀

  • Allen Sheppard's profile picture

    @ajshpprd

    @liquidai For something that just sits there watching all day I'd take the slow lane at 37W every time. Speed only matters when someone's waiting on it