Models / gpt_oss

gpt-oss-20b

Reconstructed from its own config.json with no weights read. 6.4M downloads on Hugging Face.

Our count against the checkpoint

The left number comes from the graph. The right one is the number of scalars in the published weight files. Nothing on this page was tuned to make them agree.

Derived from structure

20.91B

20,908,128,448 parameters

In the published checkpoint

20.91B

20,914,757,184 scalars · safetensors.total, read 2026-09-06

Delta

-0.03%

quantized This checkpoint is stored quantized (mxfp4). The tensor count in the file counts stored elements under a packing scheme, not logical parameters, so the two numbers below are not measuring the same thing in either direction.

What it costs to run

Cost is a roofline estimate on the priced GPU for 10 epochs at batch 32 over 50,000 samples (assumed; no dataset attached). GPU fit is fp32 weights plus gradients plus two Adam moments (16 bytes per parameter) with 1.3x headroom; activations are not included and grow with batch size.

Layers

148

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$80906.55

Card	Memory
T4 (16GB)	weights + activations	does not fit
A100 (40GB)	weights + activations	does not fit
H100 (80GB)	weights + activations	does not fit

Structure

150 nodes. Output shapes are propagated from the input shape, batch dimension excluded.

	Layer	Type	Output shape
1	Input	Input	1 × 131072
2	Embedding	Embedding	1 × 131072 × 2880
3	RoPE	RoPE	1 × 131072 × 2880
4	RMSNorm_1_1	RMSNorm	1 × 131072 × 2880
5	Attention_1	Grouped Query Attn	1 × 131072 × 2880
6	Add_1_attn	Add	1 × 131072 × 2880
7	RMSNorm_1_2	RMSNorm	1 × 131072 × 2880
8	MoE_1	MoE Layer	1 × 131072 × 2880
9	Add_1_ffn	Add	1 × 131072 × 2880
10	RMSNorm_2_1	RMSNorm	1 × 131072 × 2880
11	Attention_2	Grouped Query Attn	1 × 131072 × 2880
12	Add_2_attn	Add	1 × 131072 × 2880
13	RMSNorm_2_2	RMSNorm	1 × 131072 × 2880
14	MoE_2	MoE Layer	1 × 131072 × 2880
15	Add_2_ffn	Add	1 × 131072 × 2880
16	RMSNorm_3_1	RMSNorm	1 × 131072 × 2880
17	Attention_3	Grouped Query Attn	1 × 131072 × 2880
18	Add_3_attn	Add	1 × 131072 × 2880
19	RMSNorm_3_2	RMSNorm	1 × 131072 × 2880
20	MoE_3	MoE Layer	1 × 131072 × 2880
21	Add_3_ffn	Add	1 × 131072 × 2880
22	RMSNorm_4_1	RMSNorm	1 × 131072 × 2880
23	Attention_4	Grouped Query Attn	1 × 131072 × 2880
24	Add_4_attn	Add	1 × 131072 × 2880
25	RMSNorm_4_2	RMSNorm	1 × 131072 × 2880
26	MoE_4	MoE Layer	1 × 131072 × 2880
27	Add_4_ffn	Add	1 × 131072 × 2880
28	RMSNorm_5_1	RMSNorm	1 × 131072 × 2880
29	Attention_5	Grouped Query Attn	1 × 131072 × 2880
30	Add_5_attn	Add	1 × 131072 × 2880
31	RMSNorm_5_2	RMSNorm	1 × 131072 × 2880
32	MoE_5	MoE Layer	1 × 131072 × 2880
33	Add_5_ffn	Add	1 × 131072 × 2880
34	RMSNorm_6_1	RMSNorm	1 × 131072 × 2880
35	Attention_6	Grouped Query Attn	1 × 131072 × 2880
36	Add_6_attn	Add	1 × 131072 × 2880
37	RMSNorm_6_2	RMSNorm	1 × 131072 × 2880
38	MoE_6	MoE Layer	1 × 131072 × 2880
39	Add_6_ffn	Add	1 × 131072 × 2880
40	RMSNorm_7_1	RMSNorm	1 × 131072 × 2880
41	Attention_7	Grouped Query Attn	1 × 131072 × 2880
42	Add_7_attn	Add	1 × 131072 × 2880
43	RMSNorm_7_2	RMSNorm	1 × 131072 × 2880
44	MoE_7	MoE Layer	1 × 131072 × 2880
45	Add_7_ffn	Add	1 × 131072 × 2880
46	RMSNorm_8_1	RMSNorm	1 × 131072 × 2880
47	Attention_8	Grouped Query Attn	1 × 131072 × 2880
48	Add_8_attn	Add	1 × 131072 × 2880
49	RMSNorm_8_2	RMSNorm	1 × 131072 × 2880
50	MoE_8	MoE Layer	1 × 131072 × 2880
51	Add_8_ffn	Add	1 × 131072 × 2880
52	RMSNorm_9_1	RMSNorm	1 × 131072 × 2880
53	Attention_9	Grouped Query Attn	1 × 131072 × 2880
54	Add_9_attn	Add	1 × 131072 × 2880
55	RMSNorm_9_2	RMSNorm	1 × 131072 × 2880
56	MoE_9	MoE Layer	1 × 131072 × 2880
57	Add_9_ffn	Add	1 × 131072 × 2880
58	RMSNorm_10_1	RMSNorm	1 × 131072 × 2880
59	Attention_10	Grouped Query Attn	1 × 131072 × 2880
60	Add_10_attn	Add	1 × 131072 × 2880
61	RMSNorm_10_2	RMSNorm	1 × 131072 × 2880
62	MoE_10	MoE Layer	1 × 131072 × 2880
63	Add_10_ffn	Add	1 × 131072 × 2880
64	RMSNorm_11_1	RMSNorm	1 × 131072 × 2880
65	Attention_11	Grouped Query Attn	1 × 131072 × 2880
66	Add_11_attn	Add	1 × 131072 × 2880
67	RMSNorm_11_2	RMSNorm	1 × 131072 × 2880
68	MoE_11	MoE Layer	1 × 131072 × 2880
69	Add_11_ffn	Add	1 × 131072 × 2880
70	RMSNorm_12_1	RMSNorm	1 × 131072 × 2880
71	Attention_12	Grouped Query Attn	1 × 131072 × 2880
72	Add_12_attn	Add	1 × 131072 × 2880
73	RMSNorm_12_2	RMSNorm	1 × 131072 × 2880
74	MoE_12	MoE Layer	1 × 131072 × 2880
75	Add_12_ffn	Add	1 × 131072 × 2880
76	RMSNorm_13_1	RMSNorm	1 × 131072 × 2880
77	Attention_13	Grouped Query Attn	1 × 131072 × 2880
78	Add_13_attn	Add	1 × 131072 × 2880
79	RMSNorm_13_2	RMSNorm	1 × 131072 × 2880
80	MoE_13	MoE Layer	1 × 131072 × 2880
81	Add_13_ffn	Add	1 × 131072 × 2880
82	RMSNorm_14_1	RMSNorm	1 × 131072 × 2880
83	Attention_14	Grouped Query Attn	1 × 131072 × 2880
84	Add_14_attn	Add	1 × 131072 × 2880
85	RMSNorm_14_2	RMSNorm	1 × 131072 × 2880
86	MoE_14	MoE Layer	1 × 131072 × 2880
87	Add_14_ffn	Add	1 × 131072 × 2880
88	RMSNorm_15_1	RMSNorm	1 × 131072 × 2880
89	Attention_15	Grouped Query Attn	1 × 131072 × 2880
90	Add_15_attn	Add	1 × 131072 × 2880
91	RMSNorm_15_2	RMSNorm	1 × 131072 × 2880
92	MoE_15	MoE Layer	1 × 131072 × 2880
93	Add_15_ffn	Add	1 × 131072 × 2880
94	RMSNorm_16_1	RMSNorm	1 × 131072 × 2880
95	Attention_16	Grouped Query Attn	1 × 131072 × 2880
96	Add_16_attn	Add	1 × 131072 × 2880
97	RMSNorm_16_2	RMSNorm	1 × 131072 × 2880
98	MoE_16	MoE Layer	1 × 131072 × 2880
99	Add_16_ffn	Add	1 × 131072 × 2880
100	RMSNorm_17_1	RMSNorm	1 × 131072 × 2880
101	Attention_17	Grouped Query Attn	1 × 131072 × 2880
102	Add_17_attn	Add	1 × 131072 × 2880
103	RMSNorm_17_2	RMSNorm	1 × 131072 × 2880
104	MoE_17	MoE Layer	1 × 131072 × 2880
105	Add_17_ffn	Add	1 × 131072 × 2880
106	RMSNorm_18_1	RMSNorm	1 × 131072 × 2880
107	Attention_18	Grouped Query Attn	1 × 131072 × 2880
108	Add_18_attn	Add	1 × 131072 × 2880
109	RMSNorm_18_2	RMSNorm	1 × 131072 × 2880
110	MoE_18	MoE Layer	1 × 131072 × 2880
111	Add_18_ffn	Add	1 × 131072 × 2880
112	RMSNorm_19_1	RMSNorm	1 × 131072 × 2880
113	Attention_19	Grouped Query Attn	1 × 131072 × 2880
114	Add_19_attn	Add	1 × 131072 × 2880
115	RMSNorm_19_2	RMSNorm	1 × 131072 × 2880
116	MoE_19	MoE Layer	1 × 131072 × 2880
117	Add_19_ffn	Add	1 × 131072 × 2880
118	RMSNorm_20_1	RMSNorm	1 × 131072 × 2880
119	Attention_20	Grouped Query Attn	1 × 131072 × 2880
120	Add_20_attn	Add	1 × 131072 × 2880
121	RMSNorm_20_2	RMSNorm	1 × 131072 × 2880
122	MoE_20	MoE Layer	1 × 131072 × 2880
123	Add_20_ffn	Add	1 × 131072 × 2880
124	RMSNorm_21_1	RMSNorm	1 × 131072 × 2880
125	Attention_21	Grouped Query Attn	1 × 131072 × 2880
126	Add_21_attn	Add	1 × 131072 × 2880
127	RMSNorm_21_2	RMSNorm	1 × 131072 × 2880
128	MoE_21	MoE Layer	1 × 131072 × 2880
129	Add_21_ffn	Add	1 × 131072 × 2880
130	RMSNorm_22_1	RMSNorm	1 × 131072 × 2880
131	Attention_22	Grouped Query Attn	1 × 131072 × 2880
132	Add_22_attn	Add	1 × 131072 × 2880
133	RMSNorm_22_2	RMSNorm	1 × 131072 × 2880
134	MoE_22	MoE Layer	1 × 131072 × 2880
135	Add_22_ffn	Add	1 × 131072 × 2880
136	RMSNorm_23_1	RMSNorm	1 × 131072 × 2880
137	Attention_23	Grouped Query Attn	1 × 131072 × 2880
138	Add_23_attn	Add	1 × 131072 × 2880
139	RMSNorm_23_2	RMSNorm	1 × 131072 × 2880
140	MoE_23	MoE Layer	1 × 131072 × 2880
141	Add_23_ffn	Add	1 × 131072 × 2880
142	RMSNorm_24_1	RMSNorm	1 × 131072 × 2880
143	Attention_24	Grouped Query Attn	1 × 131072 × 2880
144	Add_24_attn	Add	1 × 131072 × 2880
145	RMSNorm_24_2	RMSNorm	1 × 131072 × 2880
146	MoE_24	MoE Layer	1 × 131072 × 2880
147	Add_24_ffn	Add	1 × 131072 × 2880
148	Final_RMSNorm	RMSNorm	1 × 131072 × 2880
149	LM_Head	Linear	1 × 131072 × 201088
150	Output	Output	1 × 131072 × 201088

What the verifier says

infoMoE layers require an auxiliary router z-loss + load-balance loss during training to prevent expert collapse. This is not visible in the architecture diagram but must be in the training loop. Applies to all 24: MoE_1, MoE_2, MoE_3, MoE_4, MoE_5, MoE_6, MoE_7, MoE_8, MoE_9, MoE_10, MoE_11, MoE_12, MoE_13, MoE_14, MoE_15, MoE_16, MoE_17, MoE_18, MoE_19, MoE_20, MoE_21, MoE_22, MoE_23, MoE_24. Fix: Add a note on these layers. Typical aux_loss coefficient: 1e-2 (Mixtral/Switch Transformer).
moe-no-aux-loss

infoAt 24 stacked attention layers, residual-branch outputs add up; unscaled init lets activation variance grow with depth. GPT-2/LLaMA-family models scale the residual projections by depth (N(0, 0.02 / √(2L))). Fix: Scale residual output projections by depth: nn.init.normal_(w, std=0.02 / math.sqrt(2 * n_layers))
deep-attention-default-init

Do this to your own model

Same numbers, on a model in your repo, in one command. No account.

pip install neurarch-trace
neurarch-trace openai/gpt-oss-20b --plan --share

Paste a model instead Markdown twin Index JSON POST /api/v1/plan