- It seems like it’ll be the best local model that can be ran fast if you have a lot of RAM and medium VRAM.
- It uses a shared expert (like deepseek and llama4) so it’ll be even faster on partial offloaded setups.
- There is a ton of options for fine tuning or training from one of their many partially trainined checkpoints.
- I’m hoping for a good reasoning finetune. Hoping Nous does it.
- It has a unique voice because it has very little synthetic data in it.
llama.CPP support is in the works, and hopefully won’t take too long since it’s architecture is reused from other models llamacpp already supports.
Are y’all as excited as I am? Also is there any other upcoming release that you’re excited for?
Yes with llamacpp its easy to put just the experts on the CPU. Since only some of the experts are used every time, the GB moved to RAM slows things down way less than moving parts of the model that are used every time. And now parts that are used every time get to stay on the GPU. I was able to get llama4 scout running at around 15 T/s on 96GB RAM and 24GB VRAM with a large context. The whole GGUF was about 80GB.
Also they actually are a Chinese company. I am pretty sure it is the company that makes RedNote (Chinese tiktok) and thats why they had access to so much non-synthetic data. I tried the demo on huggingface and never got any Chinese characters.
I also really enjoyed it’s prose. I think this will be a winner for creative writing.