This is a very interesting strategy that might pay off. This model is a very good option for enterprise self host. I would argue a lot of companies are VRAM constrained rather than compute constrained. You could fit 4-5 running instances on one H100 cluster where you can only fit 1-2 Kimi K2 or GLM5.

▲

2001zhaozhao 6 hours ago | parent | next [-]

This is 128B dense though. the K/V cache on long context is going to be massive

▲

Havoc 5 hours ago | parent | next [-]

Don’t think kv size correlates to dense/moe

	▲	zozbot234 5 hours ago \| parent [-]
		KV size correlates with attention parameters which are a subset of active parameters. So a typical MoE model will have way lower KV size than a dense model of equal total parameter count.

▲

syntaxing 4 hours ago | parent | prev [-]

With turbo quant, you would reduce it by over 6X.

▲

sayYayToLife 5 hours ago | parent | prev [-]

[dead]