this post was submitted on 07 Apr 2025
22 points (89.3% liked)

LocalLLaMA

2878 readers
34 users here now

Welcome to LocalLLaMA! Here we discuss running and developing machine learning models at home. Lets explore cutting edge open source neural network technology together.

Get support from the community! Ask questions, share prompts, discuss benchmarks, get hyped at the latest and greatest model releases! Enjoy talking about our awesome hobby.

As ambassadors of the self-hosting machine learning community, we strive to support each other and share our enthusiasm in a positive constructive way.

founded 2 years ago
MODERATORS
 

General consensus seems to be that llama4 was a flop. A head of meta AI research division was let go.

Do you think it was a bad fp32 conversion, or just unerwhelming models all around?

2t parameters was a big increase without much gain. If throwing compute and parameters isnt working to stay competitive anymore, how do you think the next big performance gains will be made? Better CoT reasoning patterns? Omnimodal? something entirely new?

top 6 comments
sorted by: hot top controversial new old
[–] [email protected] 7 points 2 weeks ago (1 children)

I think the next bit of performance may be leaning hard into QAT. We know there is a lot of wasted precision in models, so the more we understand that during training the better quality small quants can get.

I also think diffusion LLMs ability to change previous tokens is amazing. As well as the ability to iteratively use an auto regressive LLM to increase output quality.

I think a mix of QAT and iterative interference will bring the biggest upgrades to local use. It'll give you a smaller higher quality model thay you can decide to run for even longer for higher quality outputs.

[–] [email protected] 3 points 2 weeks ago (1 children)

Hmmn, never heard of QAT. What does it stand for?

[–] [email protected] 4 points 1 week ago (1 children)

https://pytorch.org/blog/quantization-aware-training/

I had heard of it but I'm not aware of public models implementing this

[–] [email protected] 3 points 1 week ago (1 children)

Here is link for ollama for Gemma 3 QAT https://ollama.com/eramax/gemma-3-27b-it-qat:q4_0

There are ggufs around if you want to try it on another back end.

[–] [email protected] 1 points 1 week ago

Thanks. I'll try it out!

[–] [email protected] 4 points 2 weeks ago

I'm pretty convinced it is just a bad set of models.

I'm waiting for qwq2 or something. Llama hasn't been my choice for a while.