As the competition for the best LLM doesn’t stop for a second, as soon as Kimi-K2 made splashes by being as good as some frontier private models, Qwen comes back with a revised model on their own: Behold Qwen3-235B-A22B-Instruct-2507! (those names are getting out of hand already, but it’s still better than whatever openAI calls their stuff) Both Qwen and Moonshot, the creator of Kimi, are backed by Alibaba, the e-commerce giant. The biggest advantage of Qwen is that it’s a fifth of the size of Kimi’s, which has a trillion parameters, while Qwen has 235 billion. This makes it way cheaper to deploy. Qwen has an Apache license, and Kimi, a modified MIT one. Another nice thing about the new qwen is that they now have separate Thinking and non- thinking models, which makes it much easier to use with agents, for example. I was trying to get Kimi to run on Blablador, but that would be ridiculously expensive, given that their minimum recommended hardware is 16x H200 with 141gb of vram each, which would be something like 6 JUPITER nodes, just for a single model. Makes little sense. So, I got the latest Qwen, released last night, running temporarily here! Happy barking!
