GPT-OSS

openai MOE text Apache-2.0 2025-08-12

架构

Total params
120 B
Active params
5 B
Layers
36
Context
128 k

详细规格

Hidden size
2880
FFN size
11520
Attention heads
32
KV heads
8
Head dim
90
Vocab size
200000
Attention type
gqa
MoE experts
128
MoE top-k
4
Expert hidden
2880

算子拆解 (per token)

算子 FLOPs / token Bytes / token
matmul 7.17e+9 7.17e+9
attention 2.04e+9 2.81e+9
moe-gate 1.33e+7 2.39e+9
rmsnorm 1.04e+6 4.15e+5

兼容硬件