https://huggingface.co/openbmb/MiniCPM5-1B

What even is this thing? MiniCPM 4.6 was a tuned Qwen 3.5 0.8B, but this looks like something else. It doesn't have vision, and it apparently has its own tokenizer. The model itself is aware of existence of Qwen 2.5, but says it's not that. Is it a new model from scratch?

I don't use agents, but I checked out mradermacher's Heretic Q6_K a bit and it seems to work quite fine. Pretty reasonable and brief thinking, unlike the "but wait" infinite loop of newer Qwens. And its speech pattern seems different from other small models I've tried.

Hey, does nobody here get hyped about new tiny models anymore? Where's everybody?