Lumimaid-Nexus-2x12B-MoE-vllm

Lumimaid-Nexus-2x12B-MoE-vllm repackages mrcuddle’s existing mixture-of-experts merge for the classic Mixtral tensor layout. Its author says numerical weight values remain unchanged. The source combines NeverSleep/Lumimaid-v0.2-12B and ReadyArt/Dark-Nexus-12B-v2.0, with narrative consistency and character interaction described as their respective focuses.

The conversion repository was created October 14, while the original merge repository dates to October 8. Both cards declare Apache-2.0. Downloadable BF16 shards require an inference engine and suitable hardware; the card names vLLM, llama.cpp converters and older Transformers as intended consumers. AITrending has not run these engines against the checkpoint, measured memory requirements or independently compared converted weight values.