RealFeedApp beta
- Qwen 3.6 35B A4B+ tested with expanded MoE
The new experimental llama.cpp feature expands MoE expert activation to 16, 24, or all available experts, replacing the fixed Top-K limit with a tunable compute budget. This allows activating more experts than in the standard implementation, but requires more memory and increases accelerator load.
- Field
- applying language models
- What they did
- Developer vagrillo added an experimental feature to the llama.cpp library that allows activating more experts in MoE models during runtime than permitted during standard training.
- Why it matters
- This could improve performance on complex tasks by utilizing more of the model's knowledge, though it requires more computational resources and memory.
#llama.cpp#moe#inference#qwen#vllm
