Hi!
In the new versions of DeepSeek they introduced an enhancement to the residual connections named manifold constrained hyper connections
https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash
The model is 284b size and it's may be easier to launch than Kimi, currently
I wonder how it will match the current abliteration framework or if new inventions are needed
Thank you!
Hi!
In the new versions of DeepSeek they introduced an enhancement to the residual connections named manifold constrained hyper connections
https://huggingface.co/deepseek-ai/DeepSeek-V4-Flash
The model is 284b size and it's may be easier to launch than Kimi, currently
I wonder how it will match the current abliteration framework or if new inventions are needed
Thank you!