LongCat 2.0 NVFP4

Target checkpoint with ModelOpt NVFP4 base weights and 16 BF16 over-embedding tables. The weight index references 173 base shards and 16 OE shards.

Attention/projection and over-embedding data retain their checkpoint dtypes; NVFP4 values are accompanied by their quantization scales. The tokenizer, chat template, generation settings and quantization configuration are included.

The target uses LongcatFlashForCausalLM / longcat_flash, with 38 paired layers. Use a runtime implementing this LongCat architecture and NVFP4 format; the validated runtime is the TokenSpeed LongCat implementation.

Downloads last month
118
Safetensors
Model size
758B params
Tensor type
BF16
ยท
U8
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support