Models / qwen3_moe

Qwen3-30B-A3B

Reconstructed from its own config.json with no weights read. 2.3M 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

30.53B

30,532,262,272 parameters

In the published checkpoint

30.53B

30,532,122,624 scalars · safetensors.total, read 2026-09-06

Delta

+0.00%

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

292

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$17009.39

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

Structure

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

	Layer	Type	Output shape
1	Input	Input	1 × 40960
2	Embedding	Embedding	1 × 40960 × 2048
3	RoPE	RoPE	1 × 40960 × 2048
4	RMSNorm_1_1	RMSNorm	1 × 40960 × 2048
5	Attention_1	Grouped Query Attn	1 × 40960 × 2048
6	Add_1_attn	Add	1 × 40960 × 2048
7	RMSNorm_1_2	RMSNorm	1 × 40960 × 2048
8	MoE_1	MoE Layer	1 × 40960 × 2048
9	Add_1_ffn	Add	1 × 40960 × 2048
10	RMSNorm_2_1	RMSNorm	1 × 40960 × 2048
11	Attention_2	Grouped Query Attn	1 × 40960 × 2048
12	Add_2_attn	Add	1 × 40960 × 2048
13	RMSNorm_2_2	RMSNorm	1 × 40960 × 2048
14	MoE_2	MoE Layer	1 × 40960 × 2048
15	Add_2_ffn	Add	1 × 40960 × 2048
16	RMSNorm_3_1	RMSNorm	1 × 40960 × 2048
17	Attention_3	Grouped Query Attn	1 × 40960 × 2048
18	Add_3_attn	Add	1 × 40960 × 2048
19	RMSNorm_3_2	RMSNorm	1 × 40960 × 2048
20	MoE_3	MoE Layer	1 × 40960 × 2048
21	Add_3_ffn	Add	1 × 40960 × 2048
22	RMSNorm_4_1	RMSNorm	1 × 40960 × 2048
23	Attention_4	Grouped Query Attn	1 × 40960 × 2048
24	Add_4_attn	Add	1 × 40960 × 2048
25	RMSNorm_4_2	RMSNorm	1 × 40960 × 2048
26	MoE_4	MoE Layer	1 × 40960 × 2048
27	Add_4_ffn	Add	1 × 40960 × 2048
28	RMSNorm_5_1	RMSNorm	1 × 40960 × 2048
29	Attention_5	Grouped Query Attn	1 × 40960 × 2048
30	Add_5_attn	Add	1 × 40960 × 2048
31	RMSNorm_5_2	RMSNorm	1 × 40960 × 2048
32	MoE_5	MoE Layer	1 × 40960 × 2048
33	Add_5_ffn	Add	1 × 40960 × 2048
34	RMSNorm_6_1	RMSNorm	1 × 40960 × 2048
35	Attention_6	Grouped Query Attn	1 × 40960 × 2048
36	Add_6_attn	Add	1 × 40960 × 2048
37	RMSNorm_6_2	RMSNorm	1 × 40960 × 2048
38	MoE_6	MoE Layer	1 × 40960 × 2048
39	Add_6_ffn	Add	1 × 40960 × 2048
40	RMSNorm_7_1	RMSNorm	1 × 40960 × 2048
41	Attention_7	Grouped Query Attn	1 × 40960 × 2048
42	Add_7_attn	Add	1 × 40960 × 2048
43	RMSNorm_7_2	RMSNorm	1 × 40960 × 2048
44	MoE_7	MoE Layer	1 × 40960 × 2048
45	Add_7_ffn	Add	1 × 40960 × 2048
46	RMSNorm_8_1	RMSNorm	1 × 40960 × 2048
47	Attention_8	Grouped Query Attn	1 × 40960 × 2048
48	Add_8_attn	Add	1 × 40960 × 2048
49	RMSNorm_8_2	RMSNorm	1 × 40960 × 2048
50	MoE_8	MoE Layer	1 × 40960 × 2048
51	Add_8_ffn	Add	1 × 40960 × 2048
52	RMSNorm_9_1	RMSNorm	1 × 40960 × 2048
53	Attention_9	Grouped Query Attn	1 × 40960 × 2048
54	Add_9_attn	Add	1 × 40960 × 2048
55	RMSNorm_9_2	RMSNorm	1 × 40960 × 2048
56	MoE_9	MoE Layer	1 × 40960 × 2048
57	Add_9_ffn	Add	1 × 40960 × 2048
58	RMSNorm_10_1	RMSNorm	1 × 40960 × 2048
59	Attention_10	Grouped Query Attn	1 × 40960 × 2048
60	Add_10_attn	Add	1 × 40960 × 2048
61	RMSNorm_10_2	RMSNorm	1 × 40960 × 2048
62	MoE_10	MoE Layer	1 × 40960 × 2048
63	Add_10_ffn	Add	1 × 40960 × 2048
64	RMSNorm_11_1	RMSNorm	1 × 40960 × 2048
65	Attention_11	Grouped Query Attn	1 × 40960 × 2048
66	Add_11_attn	Add	1 × 40960 × 2048
67	RMSNorm_11_2	RMSNorm	1 × 40960 × 2048
68	MoE_11	MoE Layer	1 × 40960 × 2048
69	Add_11_ffn	Add	1 × 40960 × 2048
70	RMSNorm_12_1	RMSNorm	1 × 40960 × 2048
71	Attention_12	Grouped Query Attn	1 × 40960 × 2048
72	Add_12_attn	Add	1 × 40960 × 2048
73	RMSNorm_12_2	RMSNorm	1 × 40960 × 2048
74	MoE_12	MoE Layer	1 × 40960 × 2048
75	Add_12_ffn	Add	1 × 40960 × 2048
76	RMSNorm_13_1	RMSNorm	1 × 40960 × 2048
77	Attention_13	Grouped Query Attn	1 × 40960 × 2048
78	Add_13_attn	Add	1 × 40960 × 2048
79	RMSNorm_13_2	RMSNorm	1 × 40960 × 2048
80	MoE_13	MoE Layer	1 × 40960 × 2048
81	Add_13_ffn	Add	1 × 40960 × 2048
82	RMSNorm_14_1	RMSNorm	1 × 40960 × 2048
83	Attention_14	Grouped Query Attn	1 × 40960 × 2048
84	Add_14_attn	Add	1 × 40960 × 2048
85	RMSNorm_14_2	RMSNorm	1 × 40960 × 2048
86	MoE_14	MoE Layer	1 × 40960 × 2048
87	Add_14_ffn	Add	1 × 40960 × 2048
88	RMSNorm_15_1	RMSNorm	1 × 40960 × 2048
89	Attention_15	Grouped Query Attn	1 × 40960 × 2048
90	Add_15_attn	Add	1 × 40960 × 2048
91	RMSNorm_15_2	RMSNorm	1 × 40960 × 2048
92	MoE_15	MoE Layer	1 × 40960 × 2048
93	Add_15_ffn	Add	1 × 40960 × 2048
94	RMSNorm_16_1	RMSNorm	1 × 40960 × 2048
95	Attention_16	Grouped Query Attn	1 × 40960 × 2048
96	Add_16_attn	Add	1 × 40960 × 2048
97	RMSNorm_16_2	RMSNorm	1 × 40960 × 2048
98	MoE_16	MoE Layer	1 × 40960 × 2048
99	Add_16_ffn	Add	1 × 40960 × 2048
100	RMSNorm_17_1	RMSNorm	1 × 40960 × 2048
101	Attention_17	Grouped Query Attn	1 × 40960 × 2048
102	Add_17_attn	Add	1 × 40960 × 2048
103	RMSNorm_17_2	RMSNorm	1 × 40960 × 2048
104	MoE_17	MoE Layer	1 × 40960 × 2048
105	Add_17_ffn	Add	1 × 40960 × 2048
106	RMSNorm_18_1	RMSNorm	1 × 40960 × 2048
107	Attention_18	Grouped Query Attn	1 × 40960 × 2048
108	Add_18_attn	Add	1 × 40960 × 2048
109	RMSNorm_18_2	RMSNorm	1 × 40960 × 2048
110	MoE_18	MoE Layer	1 × 40960 × 2048
111	Add_18_ffn	Add	1 × 40960 × 2048
112	RMSNorm_19_1	RMSNorm	1 × 40960 × 2048
113	Attention_19	Grouped Query Attn	1 × 40960 × 2048
114	Add_19_attn	Add	1 × 40960 × 2048
115	RMSNorm_19_2	RMSNorm	1 × 40960 × 2048
116	MoE_19	MoE Layer	1 × 40960 × 2048
117	Add_19_ffn	Add	1 × 40960 × 2048
118	RMSNorm_20_1	RMSNorm	1 × 40960 × 2048
119	Attention_20	Grouped Query Attn	1 × 40960 × 2048
120	Add_20_attn	Add	1 × 40960 × 2048
121	RMSNorm_20_2	RMSNorm	1 × 40960 × 2048
122	MoE_20	MoE Layer	1 × 40960 × 2048
123	Add_20_ffn	Add	1 × 40960 × 2048
124	RMSNorm_21_1	RMSNorm	1 × 40960 × 2048
125	Attention_21	Grouped Query Attn	1 × 40960 × 2048
126	Add_21_attn	Add	1 × 40960 × 2048
127	RMSNorm_21_2	RMSNorm	1 × 40960 × 2048
128	MoE_21	MoE Layer	1 × 40960 × 2048
129	Add_21_ffn	Add	1 × 40960 × 2048
130	RMSNorm_22_1	RMSNorm	1 × 40960 × 2048
131	Attention_22	Grouped Query Attn	1 × 40960 × 2048
132	Add_22_attn	Add	1 × 40960 × 2048
133	RMSNorm_22_2	RMSNorm	1 × 40960 × 2048
134	MoE_22	MoE Layer	1 × 40960 × 2048
135	Add_22_ffn	Add	1 × 40960 × 2048
136	RMSNorm_23_1	RMSNorm	1 × 40960 × 2048
137	Attention_23	Grouped Query Attn	1 × 40960 × 2048
138	Add_23_attn	Add	1 × 40960 × 2048
139	RMSNorm_23_2	RMSNorm	1 × 40960 × 2048
140	MoE_23	MoE Layer	1 × 40960 × 2048
141	Add_23_ffn	Add	1 × 40960 × 2048
142	RMSNorm_24_1	RMSNorm	1 × 40960 × 2048
143	Attention_24	Grouped Query Attn	1 × 40960 × 2048
144	Add_24_attn	Add	1 × 40960 × 2048
145	RMSNorm_24_2	RMSNorm	1 × 40960 × 2048
146	MoE_24	MoE Layer	1 × 40960 × 2048
147	Add_24_ffn	Add	1 × 40960 × 2048
148	RMSNorm_25_1	RMSNorm	1 × 40960 × 2048
149	Attention_25	Grouped Query Attn	1 × 40960 × 2048
150	Add_25_attn	Add	1 × 40960 × 2048
151	RMSNorm_25_2	RMSNorm	1 × 40960 × 2048
152	MoE_25	MoE Layer	1 × 40960 × 2048
153	Add_25_ffn	Add	1 × 40960 × 2048
154	RMSNorm_26_1	RMSNorm	1 × 40960 × 2048
155	Attention_26	Grouped Query Attn	1 × 40960 × 2048
156	Add_26_attn	Add	1 × 40960 × 2048
157	RMSNorm_26_2	RMSNorm	1 × 40960 × 2048
158	MoE_26	MoE Layer	1 × 40960 × 2048
159	Add_26_ffn	Add	1 × 40960 × 2048
160	RMSNorm_27_1	RMSNorm	1 × 40960 × 2048
161	Attention_27	Grouped Query Attn	1 × 40960 × 2048
162	Add_27_attn	Add	1 × 40960 × 2048
163	RMSNorm_27_2	RMSNorm	1 × 40960 × 2048
164	MoE_27	MoE Layer	1 × 40960 × 2048
165	Add_27_ffn	Add	1 × 40960 × 2048
166	RMSNorm_28_1	RMSNorm	1 × 40960 × 2048
167	Attention_28	Grouped Query Attn	1 × 40960 × 2048
168	Add_28_attn	Add	1 × 40960 × 2048
169	RMSNorm_28_2	RMSNorm	1 × 40960 × 2048
170	MoE_28	MoE Layer	1 × 40960 × 2048
171	Add_28_ffn	Add	1 × 40960 × 2048
172	RMSNorm_29_1	RMSNorm	1 × 40960 × 2048
173	Attention_29	Grouped Query Attn	1 × 40960 × 2048
174	Add_29_attn	Add	1 × 40960 × 2048
175	RMSNorm_29_2	RMSNorm	1 × 40960 × 2048
176	MoE_29	MoE Layer	1 × 40960 × 2048
177	Add_29_ffn	Add	1 × 40960 × 2048
178	RMSNorm_30_1	RMSNorm	1 × 40960 × 2048
179	Attention_30	Grouped Query Attn	1 × 40960 × 2048
180	Add_30_attn	Add	1 × 40960 × 2048
181	RMSNorm_30_2	RMSNorm	1 × 40960 × 2048
182	MoE_30	MoE Layer	1 × 40960 × 2048
183	Add_30_ffn	Add	1 × 40960 × 2048
184	RMSNorm_31_1	RMSNorm	1 × 40960 × 2048
185	Attention_31	Grouped Query Attn	1 × 40960 × 2048
186	Add_31_attn	Add	1 × 40960 × 2048
187	RMSNorm_31_2	RMSNorm	1 × 40960 × 2048
188	MoE_31	MoE Layer	1 × 40960 × 2048
189	Add_31_ffn	Add	1 × 40960 × 2048
190	RMSNorm_32_1	RMSNorm	1 × 40960 × 2048
191	Attention_32	Grouped Query Attn	1 × 40960 × 2048
192	Add_32_attn	Add	1 × 40960 × 2048
193	RMSNorm_32_2	RMSNorm	1 × 40960 × 2048
194	MoE_32	MoE Layer	1 × 40960 × 2048
195	Add_32_ffn	Add	1 × 40960 × 2048
196	RMSNorm_33_1	RMSNorm	1 × 40960 × 2048
197	Attention_33	Grouped Query Attn	1 × 40960 × 2048
198	Add_33_attn	Add	1 × 40960 × 2048
199	RMSNorm_33_2	RMSNorm	1 × 40960 × 2048
200	MoE_33	MoE Layer	1 × 40960 × 2048
201	Add_33_ffn	Add	1 × 40960 × 2048
202	RMSNorm_34_1	RMSNorm	1 × 40960 × 2048
203	Attention_34	Grouped Query Attn	1 × 40960 × 2048
204	Add_34_attn	Add	1 × 40960 × 2048
205	RMSNorm_34_2	RMSNorm	1 × 40960 × 2048
206	MoE_34	MoE Layer	1 × 40960 × 2048
207	Add_34_ffn	Add	1 × 40960 × 2048
208	RMSNorm_35_1	RMSNorm	1 × 40960 × 2048
209	Attention_35	Grouped Query Attn	1 × 40960 × 2048
210	Add_35_attn	Add	1 × 40960 × 2048
211	RMSNorm_35_2	RMSNorm	1 × 40960 × 2048
212	MoE_35	MoE Layer	1 × 40960 × 2048
213	Add_35_ffn	Add	1 × 40960 × 2048
214	RMSNorm_36_1	RMSNorm	1 × 40960 × 2048
215	Attention_36	Grouped Query Attn	1 × 40960 × 2048
216	Add_36_attn	Add	1 × 40960 × 2048
217	RMSNorm_36_2	RMSNorm	1 × 40960 × 2048
218	MoE_36	MoE Layer	1 × 40960 × 2048
219	Add_36_ffn	Add	1 × 40960 × 2048
220	RMSNorm_37_1	RMSNorm	1 × 40960 × 2048
221	Attention_37	Grouped Query Attn	1 × 40960 × 2048
222	Add_37_attn	Add	1 × 40960 × 2048
223	RMSNorm_37_2	RMSNorm	1 × 40960 × 2048
224	MoE_37	MoE Layer	1 × 40960 × 2048
225	Add_37_ffn	Add	1 × 40960 × 2048
226	RMSNorm_38_1	RMSNorm	1 × 40960 × 2048
227	Attention_38	Grouped Query Attn	1 × 40960 × 2048
228	Add_38_attn	Add	1 × 40960 × 2048
229	RMSNorm_38_2	RMSNorm	1 × 40960 × 2048
230	MoE_38	MoE Layer	1 × 40960 × 2048
231	Add_38_ffn	Add	1 × 40960 × 2048
232	RMSNorm_39_1	RMSNorm	1 × 40960 × 2048
233	Attention_39	Grouped Query Attn	1 × 40960 × 2048
234	Add_39_attn	Add	1 × 40960 × 2048
235	RMSNorm_39_2	RMSNorm	1 × 40960 × 2048
236	MoE_39	MoE Layer	1 × 40960 × 2048
237	Add_39_ffn	Add	1 × 40960 × 2048
238	RMSNorm_40_1	RMSNorm	1 × 40960 × 2048
239	Attention_40	Grouped Query Attn	1 × 40960 × 2048
240	Add_40_attn	Add	1 × 40960 × 2048
241	RMSNorm_40_2	RMSNorm	1 × 40960 × 2048
242	MoE_40	MoE Layer	1 × 40960 × 2048
243	Add_40_ffn	Add	1 × 40960 × 2048
244	RMSNorm_41_1	RMSNorm	1 × 40960 × 2048
245	Attention_41	Grouped Query Attn	1 × 40960 × 2048
246	Add_41_attn	Add	1 × 40960 × 2048
247	RMSNorm_41_2	RMSNorm	1 × 40960 × 2048
248	MoE_41	MoE Layer	1 × 40960 × 2048
249	Add_41_ffn	Add	1 × 40960 × 2048
250	RMSNorm_42_1	RMSNorm	1 × 40960 × 2048
251	Attention_42	Grouped Query Attn	1 × 40960 × 2048
252	Add_42_attn	Add	1 × 40960 × 2048
253	RMSNorm_42_2	RMSNorm	1 × 40960 × 2048
254	MoE_42	MoE Layer	1 × 40960 × 2048
255	Add_42_ffn	Add	1 × 40960 × 2048
256	RMSNorm_43_1	RMSNorm	1 × 40960 × 2048
257	Attention_43	Grouped Query Attn	1 × 40960 × 2048
258	Add_43_attn	Add	1 × 40960 × 2048
259	RMSNorm_43_2	RMSNorm	1 × 40960 × 2048
260	MoE_43	MoE Layer	1 × 40960 × 2048
261	Add_43_ffn	Add	1 × 40960 × 2048
262	RMSNorm_44_1	RMSNorm	1 × 40960 × 2048
263	Attention_44	Grouped Query Attn	1 × 40960 × 2048
264	Add_44_attn	Add	1 × 40960 × 2048
265	RMSNorm_44_2	RMSNorm	1 × 40960 × 2048
266	MoE_44	MoE Layer	1 × 40960 × 2048
267	Add_44_ffn	Add	1 × 40960 × 2048
268	RMSNorm_45_1	RMSNorm	1 × 40960 × 2048
269	Attention_45	Grouped Query Attn	1 × 40960 × 2048
270	Add_45_attn	Add	1 × 40960 × 2048
271	RMSNorm_45_2	RMSNorm	1 × 40960 × 2048
272	MoE_45	MoE Layer	1 × 40960 × 2048
273	Add_45_ffn	Add	1 × 40960 × 2048
274	RMSNorm_46_1	RMSNorm	1 × 40960 × 2048
275	Attention_46	Grouped Query Attn	1 × 40960 × 2048
276	Add_46_attn	Add	1 × 40960 × 2048
277	RMSNorm_46_2	RMSNorm	1 × 40960 × 2048
278	MoE_46	MoE Layer	1 × 40960 × 2048
279	Add_46_ffn	Add	1 × 40960 × 2048
280	RMSNorm_47_1	RMSNorm	1 × 40960 × 2048
281	Attention_47	Grouped Query Attn	1 × 40960 × 2048
282	Add_47_attn	Add	1 × 40960 × 2048
283	RMSNorm_47_2	RMSNorm	1 × 40960 × 2048
284	MoE_47	MoE Layer	1 × 40960 × 2048
285	Add_47_ffn	Add	1 × 40960 × 2048
286	RMSNorm_48_1	RMSNorm	1 × 40960 × 2048
287	Attention_48	Grouped Query Attn	1 × 40960 × 2048
288	Add_48_attn	Add	1 × 40960 × 2048
289	RMSNorm_48_2	RMSNorm	1 × 40960 × 2048
290	MoE_48	MoE Layer	1 × 40960 × 2048
291	Add_48_ffn	Add	1 × 40960 × 2048
292	Final_RMSNorm	RMSNorm	1 × 40960 × 2048
293	LM_Head	Linear	1 × 40960 × 151936
294	Output	Output	1 × 40960 × 151936

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 48: 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, MoE_25, MoE_26, MoE_27, MoE_28, MoE_29, MoE_30, MoE_31, MoE_32, MoE_33, MoE_34, MoE_35, MoE_36, MoE_37, MoE_38, MoE_39, MoE_40, MoE_41, MoE_42, MoE_43, MoE_44, MoE_45, MoE_46, MoE_47, MoE_48. Fix: Add a note on these layers. Typical aux_loss coefficient: 1e-2 (Mixtral/Switch Transformer).
moe-no-aux-loss

infoAt 48 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 Qwen/Qwen3-30B-A3B --plan --share

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