hybrid-moe-30b-a3b-base

A base language model with a hybrid architecture of Mamba2, attention and mixture-of-experts layers, trained from random initialisation.

Architecture	Hybrid Mamba2 + attention + mixture-of-experts (model_type: nemotron_h)
Layers	52
Hidden size	2,688
Parameters	31.6B total, roughly 3B active per token
Initialisation	Random (trained from scratch)
Tokens seen	553,765,568,512 (553.8B)
Context length	Trained at 8,192 tokens, then continued at 32,768
Precision	BF16

This is a base model: it has had no instruction tuning or preference optimisation, so it is suited to text completion rather than chat.