# Question about trend in how AI expertise is handled > Source: > Published: 2026-08-17 04:13:03+00:00 Hi, It has been a while now that I ask myselft this question : why don’t we **specialise more small LLM** models on dedicated topics ? So we can match frontiere one in specific stuff for each one and use them in concurrency as we would do with a team of human expert collaboring together ? It has been at least three time that I saw a “kind of it” moment in AI breakthrougth, but it never really is. Is there a good reason why it won’t work out ? I mean we build Mixture of Expert whitch is essentialy leverage of sparse information of LLM based on pattern I think. Agentic whitch is sold like that, but is not. And while it works mostly, it deceive me in that regards. And now the last idea of Nvidia about “fast tuning” whitch is smart but still not that. All the three can be combine and go thurther. Just to give a better **picture** : - 1 big general MoE model as we actualy use as a general interface, connected to when it needs : - 1 math expert, - 1 physician, - 1 writter, - … wathever we need for our purpose, and they can be dense if they are small (sub 16B). Maybe an orchestrator “expert” and chain of thruth to mitigate AI internal debate (expert X know best on X, only allow 1 “naive” critic of the answer). And a scribe to consign long term memory and personalisation in files. This might also have the advantage on a multi GPU RIG to dedicate a card for a sets of model whitch elimnate the difficulty of making two card works on the same model and **mitigate exchange of big chunk of data**. It seems also that each one could have smaller optimised RAG systems for precise facts retrieval dedicated to each domain of expertise. Please link me information if you think I’m deep wrong because I misunderstood something, or if you know of similar attempts in research. A better tuned multi agent harness can already operate kind of, but models are not build with this in mind we would probably need to **fine tuned small models as a fist attempt** and make receipe for better skills. I feel that it would be usefull for humble self hosting AI to decrease cost and complexity. What are you thougth on this level1 people ? Child dream or what ? This is the topic of a very recent Level1 video, and you’re exactly right. This is the entire point of Nvidia Nemotron Lightning. It’s also the thing that makes Qwen so incredibly useful. A small model specialized can be better than a frontier model at a given task and overwhelmingly cheaper to operate. Watch the Level1 video.