It seems to me that qwen3.5 27b and 122ba10b are not too far behind the 397ba17b at least according to the benchmarks. The alibaba coding plan is selling 397ba17b for 50 dollars per month, too expensive! If say 70% of work can be done by 27b and 122ba10b, which are much easier to deploy on local PC, then releasing them will simply give people a reason to not using their coding plan. They could just use a cheaper chatgpt/claude subscription to solve the remaining harder problems.

My guess is that maybe Alibaba will gradually stop releasing powerful small models, or ensure that small models are not good enough to compete with their flagship model. Since Alibaba is one of the very few companies releasing small models, if they stop raising the bar, other companies might follow suit and slow down their progress as well. Like Z.ai, they used to release small models, but now they only release huge model and significantly increase their coding plan price (Pro plan from 30 dollars per month to 72 dollars per month).

Maybe I am too pessimistic, but I am afraid that small open source models (say below 60 GB in size) will stop evolving at some point, optimistically touch GPT-4o level. Then if you want better performance, you will either have to have hundreds of GB of VRAM to run huge local LLMs or subscribe to very expensive cloud models.