Models / minimax_m3_vl

MiniMax-M3-MXFP4

Reconstructed from its own config.json with no weights read. 122K 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

19.00B

18,999,029,760 parameters

In the published checkpoint

232B

231,876,225,920 scalars · safetensors.total, read 2026-07-17

Delta

-91.8%

quantized This checkpoint is stored quantized (quark). 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

242

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$15614498.88

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

Structure

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

	Layer	Type	Output shape
1	Input	Input	1 × 1048576
2	Embedding	Embedding	1 × 1048576 × 6144
3	Positional_Embedding	Learned Pos Embed	1 × 1048576 × 6144
4	Attention_1	Multi-Head Attention	1 × 1048576 × 6144
5	Add_1	Add	1 × 1048576 × 6144
6	LayerNorm_1_1	LayerNorm	1 × 1048576 × 6144
7	FFN_1	Feed Forward	1 × 1048576 × 6144
8	Attention_2	Multi-Head Attention	1 × 1048576 × 6144
9	Add_2	Add	1 × 1048576 × 6144
10	LayerNorm_2_1	LayerNorm	1 × 1048576 × 6144
11	FFN_2	Feed Forward	1 × 1048576 × 6144
12	Attention_3	Multi-Head Attention	1 × 1048576 × 6144
13	Add_3	Add	1 × 1048576 × 6144
14	LayerNorm_3_1	LayerNorm	1 × 1048576 × 6144
15	FFN_3	Feed Forward	1 × 1048576 × 6144
16	Attention_4	Multi-Head Attention	1 × 1048576 × 6144
17	Add_4	Add	1 × 1048576 × 6144
18	LayerNorm_4_1	LayerNorm	1 × 1048576 × 6144
19	FFN_4	Feed Forward	1 × 1048576 × 6144
20	Attention_5	Multi-Head Attention	1 × 1048576 × 6144
21	Add_5	Add	1 × 1048576 × 6144
22	LayerNorm_5_1	LayerNorm	1 × 1048576 × 6144
23	FFN_5	Feed Forward	1 × 1048576 × 6144
24	Attention_6	Multi-Head Attention	1 × 1048576 × 6144
25	Add_6	Add	1 × 1048576 × 6144
26	LayerNorm_6_1	LayerNorm	1 × 1048576 × 6144
27	FFN_6	Feed Forward	1 × 1048576 × 6144
28	Attention_7	Multi-Head Attention	1 × 1048576 × 6144
29	Add_7	Add	1 × 1048576 × 6144
30	LayerNorm_7_1	LayerNorm	1 × 1048576 × 6144
31	FFN_7	Feed Forward	1 × 1048576 × 6144
32	Attention_8	Multi-Head Attention	1 × 1048576 × 6144
33	Add_8	Add	1 × 1048576 × 6144
34	LayerNorm_8_1	LayerNorm	1 × 1048576 × 6144
35	FFN_8	Feed Forward	1 × 1048576 × 6144
36	Attention_9	Multi-Head Attention	1 × 1048576 × 6144
37	Add_9	Add	1 × 1048576 × 6144
38	LayerNorm_9_1	LayerNorm	1 × 1048576 × 6144
39	FFN_9	Feed Forward	1 × 1048576 × 6144
40	Attention_10	Multi-Head Attention	1 × 1048576 × 6144
41	Add_10	Add	1 × 1048576 × 6144
42	LayerNorm_10_1	LayerNorm	1 × 1048576 × 6144
43	FFN_10	Feed Forward	1 × 1048576 × 6144
44	Attention_11	Multi-Head Attention	1 × 1048576 × 6144
45	Add_11	Add	1 × 1048576 × 6144
46	LayerNorm_11_1	LayerNorm	1 × 1048576 × 6144
47	FFN_11	Feed Forward	1 × 1048576 × 6144
48	Attention_12	Multi-Head Attention	1 × 1048576 × 6144
49	Add_12	Add	1 × 1048576 × 6144
50	LayerNorm_12_1	LayerNorm	1 × 1048576 × 6144
51	FFN_12	Feed Forward	1 × 1048576 × 6144
52	Attention_13	Multi-Head Attention	1 × 1048576 × 6144
53	Add_13	Add	1 × 1048576 × 6144
54	LayerNorm_13_1	LayerNorm	1 × 1048576 × 6144
55	FFN_13	Feed Forward	1 × 1048576 × 6144
56	Attention_14	Multi-Head Attention	1 × 1048576 × 6144
57	Add_14	Add	1 × 1048576 × 6144
58	LayerNorm_14_1	LayerNorm	1 × 1048576 × 6144
59	FFN_14	Feed Forward	1 × 1048576 × 6144
60	Attention_15	Multi-Head Attention	1 × 1048576 × 6144
61	Add_15	Add	1 × 1048576 × 6144
62	LayerNorm_15_1	LayerNorm	1 × 1048576 × 6144
63	FFN_15	Feed Forward	1 × 1048576 × 6144
64	Attention_16	Multi-Head Attention	1 × 1048576 × 6144
65	Add_16	Add	1 × 1048576 × 6144
66	LayerNorm_16_1	LayerNorm	1 × 1048576 × 6144
67	FFN_16	Feed Forward	1 × 1048576 × 6144
68	Attention_17	Multi-Head Attention	1 × 1048576 × 6144
69	Add_17	Add	1 × 1048576 × 6144
70	LayerNorm_17_1	LayerNorm	1 × 1048576 × 6144
71	FFN_17	Feed Forward	1 × 1048576 × 6144
72	Attention_18	Multi-Head Attention	1 × 1048576 × 6144
73	Add_18	Add	1 × 1048576 × 6144
74	LayerNorm_18_1	LayerNorm	1 × 1048576 × 6144
75	FFN_18	Feed Forward	1 × 1048576 × 6144
76	Attention_19	Multi-Head Attention	1 × 1048576 × 6144
77	Add_19	Add	1 × 1048576 × 6144
78	LayerNorm_19_1	LayerNorm	1 × 1048576 × 6144
79	FFN_19	Feed Forward	1 × 1048576 × 6144
80	Attention_20	Multi-Head Attention	1 × 1048576 × 6144
81	Add_20	Add	1 × 1048576 × 6144
82	LayerNorm_20_1	LayerNorm	1 × 1048576 × 6144
83	FFN_20	Feed Forward	1 × 1048576 × 6144
84	Attention_21	Multi-Head Attention	1 × 1048576 × 6144
85	Add_21	Add	1 × 1048576 × 6144
86	LayerNorm_21_1	LayerNorm	1 × 1048576 × 6144
87	FFN_21	Feed Forward	1 × 1048576 × 6144
88	Attention_22	Multi-Head Attention	1 × 1048576 × 6144
89	Add_22	Add	1 × 1048576 × 6144
90	LayerNorm_22_1	LayerNorm	1 × 1048576 × 6144
91	FFN_22	Feed Forward	1 × 1048576 × 6144
92	Attention_23	Multi-Head Attention	1 × 1048576 × 6144
93	Add_23	Add	1 × 1048576 × 6144
94	LayerNorm_23_1	LayerNorm	1 × 1048576 × 6144
95	FFN_23	Feed Forward	1 × 1048576 × 6144
96	Attention_24	Multi-Head Attention	1 × 1048576 × 6144
97	Add_24	Add	1 × 1048576 × 6144
98	LayerNorm_24_1	LayerNorm	1 × 1048576 × 6144
99	FFN_24	Feed Forward	1 × 1048576 × 6144
100	Attention_25	Multi-Head Attention	1 × 1048576 × 6144
101	Add_25	Add	1 × 1048576 × 6144
102	LayerNorm_25_1	LayerNorm	1 × 1048576 × 6144
103	FFN_25	Feed Forward	1 × 1048576 × 6144
104	Attention_26	Multi-Head Attention	1 × 1048576 × 6144
105	Add_26	Add	1 × 1048576 × 6144
106	LayerNorm_26_1	LayerNorm	1 × 1048576 × 6144
107	FFN_26	Feed Forward	1 × 1048576 × 6144
108	Attention_27	Multi-Head Attention	1 × 1048576 × 6144
109	Add_27	Add	1 × 1048576 × 6144
110	LayerNorm_27_1	LayerNorm	1 × 1048576 × 6144
111	FFN_27	Feed Forward	1 × 1048576 × 6144
112	Attention_28	Multi-Head Attention	1 × 1048576 × 6144
113	Add_28	Add	1 × 1048576 × 6144
114	LayerNorm_28_1	LayerNorm	1 × 1048576 × 6144
115	FFN_28	Feed Forward	1 × 1048576 × 6144
116	Attention_29	Multi-Head Attention	1 × 1048576 × 6144
117	Add_29	Add	1 × 1048576 × 6144
118	LayerNorm_29_1	LayerNorm	1 × 1048576 × 6144
119	FFN_29	Feed Forward	1 × 1048576 × 6144
120	Attention_30	Multi-Head Attention	1 × 1048576 × 6144
121	Add_30	Add	1 × 1048576 × 6144
122	LayerNorm_30_1	LayerNorm	1 × 1048576 × 6144
123	FFN_30	Feed Forward	1 × 1048576 × 6144
124	Attention_31	Multi-Head Attention	1 × 1048576 × 6144
125	Add_31	Add	1 × 1048576 × 6144
126	LayerNorm_31_1	LayerNorm	1 × 1048576 × 6144
127	FFN_31	Feed Forward	1 × 1048576 × 6144
128	Attention_32	Multi-Head Attention	1 × 1048576 × 6144
129	Add_32	Add	1 × 1048576 × 6144
130	LayerNorm_32_1	LayerNorm	1 × 1048576 × 6144
131	FFN_32	Feed Forward	1 × 1048576 × 6144
132	Attention_33	Multi-Head Attention	1 × 1048576 × 6144
133	Add_33	Add	1 × 1048576 × 6144
134	LayerNorm_33_1	LayerNorm	1 × 1048576 × 6144
135	FFN_33	Feed Forward	1 × 1048576 × 6144
136	Attention_34	Multi-Head Attention	1 × 1048576 × 6144
137	Add_34	Add	1 × 1048576 × 6144
138	LayerNorm_34_1	LayerNorm	1 × 1048576 × 6144
139	FFN_34	Feed Forward	1 × 1048576 × 6144
140	Attention_35	Multi-Head Attention	1 × 1048576 × 6144
141	Add_35	Add	1 × 1048576 × 6144
142	LayerNorm_35_1	LayerNorm	1 × 1048576 × 6144
143	FFN_35	Feed Forward	1 × 1048576 × 6144
144	Attention_36	Multi-Head Attention	1 × 1048576 × 6144
145	Add_36	Add	1 × 1048576 × 6144
146	LayerNorm_36_1	LayerNorm	1 × 1048576 × 6144
147	FFN_36	Feed Forward	1 × 1048576 × 6144
148	Attention_37	Multi-Head Attention	1 × 1048576 × 6144
149	Add_37	Add	1 × 1048576 × 6144
150	LayerNorm_37_1	LayerNorm	1 × 1048576 × 6144
151	FFN_37	Feed Forward	1 × 1048576 × 6144
152	Attention_38	Multi-Head Attention	1 × 1048576 × 6144
153	Add_38	Add	1 × 1048576 × 6144
154	LayerNorm_38_1	LayerNorm	1 × 1048576 × 6144
155	FFN_38	Feed Forward	1 × 1048576 × 6144
156	Attention_39	Multi-Head Attention	1 × 1048576 × 6144
157	Add_39	Add	1 × 1048576 × 6144
158	LayerNorm_39_1	LayerNorm	1 × 1048576 × 6144
159	FFN_39	Feed Forward	1 × 1048576 × 6144
160	Attention_40	Multi-Head Attention	1 × 1048576 × 6144
161	Add_40	Add	1 × 1048576 × 6144
162	LayerNorm_40_1	LayerNorm	1 × 1048576 × 6144
163	FFN_40	Feed Forward	1 × 1048576 × 6144
164	Attention_41	Multi-Head Attention	1 × 1048576 × 6144
165	Add_41	Add	1 × 1048576 × 6144
166	LayerNorm_41_1	LayerNorm	1 × 1048576 × 6144
167	FFN_41	Feed Forward	1 × 1048576 × 6144
168	Attention_42	Multi-Head Attention	1 × 1048576 × 6144
169	Add_42	Add	1 × 1048576 × 6144
170	LayerNorm_42_1	LayerNorm	1 × 1048576 × 6144
171	FFN_42	Feed Forward	1 × 1048576 × 6144
172	Attention_43	Multi-Head Attention	1 × 1048576 × 6144
173	Add_43	Add	1 × 1048576 × 6144
174	LayerNorm_43_1	LayerNorm	1 × 1048576 × 6144
175	FFN_43	Feed Forward	1 × 1048576 × 6144
176	Attention_44	Multi-Head Attention	1 × 1048576 × 6144
177	Add_44	Add	1 × 1048576 × 6144
178	LayerNorm_44_1	LayerNorm	1 × 1048576 × 6144
179	FFN_44	Feed Forward	1 × 1048576 × 6144
180	Attention_45	Multi-Head Attention	1 × 1048576 × 6144
181	Add_45	Add	1 × 1048576 × 6144
182	LayerNorm_45_1	LayerNorm	1 × 1048576 × 6144
183	FFN_45	Feed Forward	1 × 1048576 × 6144
184	Attention_46	Multi-Head Attention	1 × 1048576 × 6144
185	Add_46	Add	1 × 1048576 × 6144
186	LayerNorm_46_1	LayerNorm	1 × 1048576 × 6144
187	FFN_46	Feed Forward	1 × 1048576 × 6144
188	Attention_47	Multi-Head Attention	1 × 1048576 × 6144
189	Add_47	Add	1 × 1048576 × 6144
190	LayerNorm_47_1	LayerNorm	1 × 1048576 × 6144
191	FFN_47	Feed Forward	1 × 1048576 × 6144
192	Attention_48	Multi-Head Attention	1 × 1048576 × 6144
193	Add_48	Add	1 × 1048576 × 6144
194	LayerNorm_48_1	LayerNorm	1 × 1048576 × 6144
195	FFN_48	Feed Forward	1 × 1048576 × 6144
196	Attention_49	Multi-Head Attention	1 × 1048576 × 6144
197	Add_49	Add	1 × 1048576 × 6144
198	LayerNorm_49_1	LayerNorm	1 × 1048576 × 6144
199	FFN_49	Feed Forward	1 × 1048576 × 6144
200	Attention_50	Multi-Head Attention	1 × 1048576 × 6144
201	Add_50	Add	1 × 1048576 × 6144
202	LayerNorm_50_1	LayerNorm	1 × 1048576 × 6144
203	FFN_50	Feed Forward	1 × 1048576 × 6144
204	Attention_51	Multi-Head Attention	1 × 1048576 × 6144
205	Add_51	Add	1 × 1048576 × 6144
206	LayerNorm_51_1	LayerNorm	1 × 1048576 × 6144
207	FFN_51	Feed Forward	1 × 1048576 × 6144
208	Attention_52	Multi-Head Attention	1 × 1048576 × 6144
209	Add_52	Add	1 × 1048576 × 6144
210	LayerNorm_52_1	LayerNorm	1 × 1048576 × 6144
211	FFN_52	Feed Forward	1 × 1048576 × 6144
212	Attention_53	Multi-Head Attention	1 × 1048576 × 6144
213	Add_53	Add	1 × 1048576 × 6144
214	LayerNorm_53_1	LayerNorm	1 × 1048576 × 6144
215	FFN_53	Feed Forward	1 × 1048576 × 6144
216	Attention_54	Multi-Head Attention	1 × 1048576 × 6144
217	Add_54	Add	1 × 1048576 × 6144
218	LayerNorm_54_1	LayerNorm	1 × 1048576 × 6144
219	FFN_54	Feed Forward	1 × 1048576 × 6144
220	Attention_55	Multi-Head Attention	1 × 1048576 × 6144
221	Add_55	Add	1 × 1048576 × 6144
222	LayerNorm_55_1	LayerNorm	1 × 1048576 × 6144
223	FFN_55	Feed Forward	1 × 1048576 × 6144
224	Attention_56	Multi-Head Attention	1 × 1048576 × 6144
225	Add_56	Add	1 × 1048576 × 6144
226	LayerNorm_56_1	LayerNorm	1 × 1048576 × 6144
227	FFN_56	Feed Forward	1 × 1048576 × 6144
228	Attention_57	Multi-Head Attention	1 × 1048576 × 6144
229	Add_57	Add	1 × 1048576 × 6144
230	LayerNorm_57_1	LayerNorm	1 × 1048576 × 6144
231	FFN_57	Feed Forward	1 × 1048576 × 6144
232	Attention_58	Multi-Head Attention	1 × 1048576 × 6144
233	Add_58	Add	1 × 1048576 × 6144
234	LayerNorm_58_1	LayerNorm	1 × 1048576 × 6144
235	FFN_58	Feed Forward	1 × 1048576 × 6144
236	Attention_59	Multi-Head Attention	1 × 1048576 × 6144
237	Add_59	Add	1 × 1048576 × 6144
238	LayerNorm_59_1	LayerNorm	1 × 1048576 × 6144
239	FFN_59	Feed Forward	1 × 1048576 × 6144
240	Attention_60	Multi-Head Attention	1 × 1048576 × 6144
241	Add_60	Add	1 × 1048576 × 6144
242	LayerNorm_60_1	LayerNorm	1 × 1048576 × 6144
243	FFN_60	Feed Forward	1 × 1048576 × 6144
244	Output	Output	1 × 1048576 × 6144

What the verifier says

info60 attention layers at embedDim 6144 cache full per-head K/V: about 1440 KB per token at fp16, which dominates memory at long context. Grouped-query attention (e.g. 8:1) would cut this ~8×; multi-head latent attention (MLA) shrinks it ~10× or more. This is the move production LLMs make; it does not change the parameter count. Fix: Switch attention to groupedQueryAttention (set numKVHeads below numHeads, e.g. numHeads/4) or mla (a low-rank cached latent).
full-mha-serving-cost

infoAt 60 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

warnAcross 60 attention layers this design caches 1440 KB per token, so a single 8,192-token sequence needs ~12.1 GB of KV cache before weights or activations. That exceeds the 4 GB budget this rule assumes for serving headroom. Fix: Cut KV width: raise the GQA ratio (fewer numKVHeads), switch to MLA, reduce depth or embedDim, or accept a shorter serving context.
kv-cache-context-budget

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 amd/MiniMax-M3-MXFP4 --plan --share

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

Other minimax_m3_vl checkpoints

MiniMax-M3

19.00B derived · -95.6% against the checkpoint

MiniMax-M3-MXFP8

19.00B derived · -95.7% against the checkpoint

MiniMax-M3-NVFP4

19.00B derived · -91.4% against the checkpoint