23 Comments

1. 

   @productgremlin

   April 7, 2025 at 12:30 pm

   They are dropping models on Saturday. We can't even have a weekend off.

   Reply
2. 

   @Roflux

   April 7, 2025 at 12:30 pm

   Massive clickbait, it's shit. Maybe try the models next time lol

   Reply
3. 

   @blender_wiki

   April 7, 2025 at 12:30 pm

   Don't speak before reading the tech spec …

   Reply
4. 

   @patrickbaumann8670

   April 7, 2025 at 12:30 pm

   but it sucks at coding sadly :^) atleast the base models, im grateful for the AI community, do your thing!

   Reply
5. 

   @grilloyoo

   April 7, 2025 at 12:30 pm

   Useless

   Reply
6. 

   @salvatorecipolla8456

   April 7, 2025 at 12:30 pm

   Inside the report they talk about reasoning:
   "The biggest challenge while post-training the Llama 4 Maverick model was maintaining a balance between multiple input modalities, reasoning, and conversational abilities."

   Reply
7. 

   @Tethysmeer

   April 7, 2025 at 12:30 pm

   would the smallest model run on a amd 395+ 128gb?

   Reply
8. 

   @surajsingh-ji5cw

   April 7, 2025 at 12:30 pm

   SUPERB INSANE CRAZY COOKED MINDBLOWING

   Reply
9. 

   @mikelaurie3850

   April 7, 2025 at 12:30 pm

   I dont think they understand infinite

   Reply
10. 

    @PaulHobbs23

    April 7, 2025 at 12:30 pm

    This model is seriously terrible. I'd expect that Meta is directly training on benchmarks at this point.

    Reply
11. 

    @splaytrees

    April 7, 2025 at 12:30 pm

    The H100, on which it runs, has 16,896 CUDA cores and 528 Tensor Cores, along with 80GB of HBM2e memory.
    (That's 44x the CUDA cores of the 730 and 40x the memory lol)

    Reply
12. 

    @cleverestx

    April 7, 2025 at 12:30 pm

    I really hope they get this model integrated with the Ray-Ban Meta glasses as the AI model it streams to and from

    Reply
13. 

    @whodidthat-p2k

    April 7, 2025 at 12:30 pm

    how are you supposed to use llama 4 if you just have a normal laptop? doesn't seem like there's a web app

    Reply
14. 

    @DarrylPolo-y9y

    April 7, 2025 at 12:30 pm

    i cant wait for the day that i can paste my 2-3000 lines of code in at once with the mess i made building it to just let it clean it up for me lol

    Reply
15. 

    @BeeRich33

    April 7, 2025 at 12:30 pm

    Big | Medium | Small. See how that works?

    Reply
16. 

    @vancass1326

    April 7, 2025 at 12:30 pm

    As long as I do not go into that conversation, I have no problem using it

    Reply
17. 

    @vancass1326

    April 7, 2025 at 12:30 pm

    When I take it off pause, it continues unabated

    Reply
18. 

    @vancass1326

    April 7, 2025 at 12:30 pm

    I crashed DeepSeek several days ago. While having it do an analysis of a design, it stuttered an went into an endless loop of the word the

    Reply
19. 

    @vancass1326

    April 7, 2025 at 12:30 pm

    I crashed DeepSeek several days ago. While having it do an analysis of a design, it stuttered an went into an endless loop of the word the

    Reply
20. 

    @DustySides

    April 7, 2025 at 12:30 pm

    Have you tried it? I downloaded and am running Maverick. It’s pure garbage. Hallucinations like Jimi Hendrix at Woodstock with full confidence. 10M context is “virtual” because it wasn’t trained on 10M token prompts. They claim it’s open source but it has commercial restrictions and naming/“Built with Llama” requirements (aka NOT open source.)

    Literally everything they’ve said to hype this is vapor. Training to pass benchmarks is easy, making it actually useful is hard. Guess which they did.

    Reply
21. 

    @XenMaximalist

    April 7, 2025 at 12:30 pm

    MIND BLOWN

    Reply
22. 

    @rumfordc

    April 7, 2025 at 12:30 pm

    it's not difficult to make the context window bigger, as big as you want even, what's difficult is making the entire context window have a meaningful effect on the output.

    Reply
23. 

    @afterthesmash

    April 7, 2025 at 12:30 pm

    The whole point of MoE is to run the table on an MoB (mixture of [synthetic] benchmarks). If the MoB isn't scoring some skill explicitly, it won't show up in the MoE, either. Poof goes everything you don't see in between the cracks. If your task lies in the groove of the easily quantified (i.e. has been already distilled into the MoB) then smooth sailing. If not, the surrounding currents likely won't take you there.

    I want a model that excels most of all in what's never been done before, and I harbour serious doubts that MoE is a magic stairway to heavenly realms as yet undiscovered.

    Be VERY WARY about excelling too directly in modelling your concerns of yesteryear.

    Reply