3 ms·
42B active params, sliding window attention. There's your tradeoff.
by moffkalast 5mo ago
42B active params, sliding window attention. There's your tradeoff.
- vlovich123 5mo agoSliding window for the draft model, not for the main. 42B for active params because it’s a sparse MoE which is a common technique for the larger models to not get bottlenecked by memory bandwidth.
- moffkalast 5mo agoSeems to be for both according to the spec [0], maybe it's wrong though. 128 sounds really tiny, I wonder if they mean some kind of blocks? [0] https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash#4-model-summary https://huggingface.co/XiaomiMiMo/MiMo-V2.5-Pro-FP4-DFlash#4...
- E-Reverance 5mo agoNo > It uses 384 routed experts (top-8) with hybrid attention (full-attention + sliding-window 128 at 6:1 ratio) over 70 layers (1 dense + 69 MoE) https://recipes.vllm.ai/XiaomiMiMo/MiMo-V2.5-Pro https://recipes.vllm.ai/XiaomiMiMo/MiMo-V2.5-Pro
- bearjaws 5mo agoGiven how "smart" some of the 26b dense models are now, I would not be surprised to see a strong 40b MoE.