
Qwen team is set to release a new model, Qwen2MoE, with reduced non-embedding parameters compared to Qwen1.5-7B. Qwen1.5-MoE-A2.7B contains only 2.0 billion non-embedding parameters, one-third of Qwen1.5-7B's size. The model aims to combine fine-tuned LLMs into a single MoE for improved AI performance on consumer-grade hardware.
If you are running consumer grade hardware, you need AI that matches. Qwen1.5-MoE-A2.7B answers this need. It has the performance of decent models (7B parameter models) but with the speed that is seen in smaller models. https://t.co/wMSAi7eQSs
Can we combine multiple fine-tuned LLMs into a single MoE? 🤔 Earlier today Qwen1.5-MoE-A2.7B a combined MoE of smaller models and probably further trained. Meta released a paper called “Branch-Train-MiX” exploring if we can combine individual “experts” (LLMs) into a single… https://t.co/Skq0OBXteo
Qwen1.5-MoE-A2.7B!!! 💡 https://t.co/e2XTjQlYff 🤗 https://t.co/vkPI73DmuZ Compared to Qwen1.5-7B, which contains 6.5 billion non-embedding parameters, Qwen1.5-MoE-A2.7B contains only 2.0 billion non-embedding parameters, approximately one-third of Qwen1.5-7B’s size. Notably,… https://t.co/C2ZlbR6BF3
