With Qwen 3 next (documentation: http://huggingface.co/collections/Qwen/qwen3-next-68c25fd6838e585db8eeea9d and test on Qwen Chat) Alibaba offers a new language model open-source with long reasoning.

Medium-sized (80B), it’s optimized for speed of response and performance, with 119 languages available.

It uses an “expert mix” (MoE) architecture that allows only the “experts” needed to handle a task to be used for more efficiency in the training phase and more speed in the response/inference phase.

via Ciprian Melian