Models / XLMRobertaModel

jina-embeddings-v3

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

567M

566,652,928 parameters

In the published checkpoint

572M

572,310,396 scalars · safetensors.total, read 2026-09-06

Delta

-0.99%

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

Will it forward-pass

Yes

Priced on

A10G (24GB)

Est. one run

$154.43

Card	Memory
T4 (16GB)	weights + activations	fits
A100 (40GB)	weights + activations	fits
H100 (80GB)	weights + activations	fits

Structure

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

	Layer	Type	Output shape
1	Input	Input	1 × 8194
2	Embedding	Embedding	1 × 8194 × 1024
3	Positional_Embedding	Learned Pos Embed	1 × 8194 × 1024
4	Attention_1	Multi-Head Attention	1 × 8194 × 1024
5	Add_1	Add	1 × 8194 × 1024
6	LayerNorm_1_1	LayerNorm	1 × 8194 × 1024
7	FFN_1	Feed Forward	1 × 8194 × 1024
8	Attention_2	Multi-Head Attention	1 × 8194 × 1024
9	Add_2	Add	1 × 8194 × 1024
10	LayerNorm_2_1	LayerNorm	1 × 8194 × 1024
11	FFN_2	Feed Forward	1 × 8194 × 1024
12	Attention_3	Multi-Head Attention	1 × 8194 × 1024
13	Add_3	Add	1 × 8194 × 1024
14	LayerNorm_3_1	LayerNorm	1 × 8194 × 1024
15	FFN_3	Feed Forward	1 × 8194 × 1024
16	Attention_4	Multi-Head Attention	1 × 8194 × 1024
17	Add_4	Add	1 × 8194 × 1024
18	LayerNorm_4_1	LayerNorm	1 × 8194 × 1024
19	FFN_4	Feed Forward	1 × 8194 × 1024
20	Attention_5	Multi-Head Attention	1 × 8194 × 1024
21	Add_5	Add	1 × 8194 × 1024
22	LayerNorm_5_1	LayerNorm	1 × 8194 × 1024
23	FFN_5	Feed Forward	1 × 8194 × 1024
24	Attention_6	Multi-Head Attention	1 × 8194 × 1024
25	Add_6	Add	1 × 8194 × 1024
26	LayerNorm_6_1	LayerNorm	1 × 8194 × 1024
27	FFN_6	Feed Forward	1 × 8194 × 1024
28	Attention_7	Multi-Head Attention	1 × 8194 × 1024
29	Add_7	Add	1 × 8194 × 1024
30	LayerNorm_7_1	LayerNorm	1 × 8194 × 1024
31	FFN_7	Feed Forward	1 × 8194 × 1024
32	Attention_8	Multi-Head Attention	1 × 8194 × 1024
33	Add_8	Add	1 × 8194 × 1024
34	LayerNorm_8_1	LayerNorm	1 × 8194 × 1024
35	FFN_8	Feed Forward	1 × 8194 × 1024
36	Attention_9	Multi-Head Attention	1 × 8194 × 1024
37	Add_9	Add	1 × 8194 × 1024
38	LayerNorm_9_1	LayerNorm	1 × 8194 × 1024
39	FFN_9	Feed Forward	1 × 8194 × 1024
40	Attention_10	Multi-Head Attention	1 × 8194 × 1024
41	Add_10	Add	1 × 8194 × 1024
42	LayerNorm_10_1	LayerNorm	1 × 8194 × 1024
43	FFN_10	Feed Forward	1 × 8194 × 1024
44	Attention_11	Multi-Head Attention	1 × 8194 × 1024
45	Add_11	Add	1 × 8194 × 1024
46	LayerNorm_11_1	LayerNorm	1 × 8194 × 1024
47	FFN_11	Feed Forward	1 × 8194 × 1024
48	Attention_12	Multi-Head Attention	1 × 8194 × 1024
49	Add_12	Add	1 × 8194 × 1024
50	LayerNorm_12_1	LayerNorm	1 × 8194 × 1024
51	FFN_12	Feed Forward	1 × 8194 × 1024
52	Attention_13	Multi-Head Attention	1 × 8194 × 1024
53	Add_13	Add	1 × 8194 × 1024
54	LayerNorm_13_1	LayerNorm	1 × 8194 × 1024
55	FFN_13	Feed Forward	1 × 8194 × 1024
56	Attention_14	Multi-Head Attention	1 × 8194 × 1024
57	Add_14	Add	1 × 8194 × 1024
58	LayerNorm_14_1	LayerNorm	1 × 8194 × 1024
59	FFN_14	Feed Forward	1 × 8194 × 1024
60	Attention_15	Multi-Head Attention	1 × 8194 × 1024
61	Add_15	Add	1 × 8194 × 1024
62	LayerNorm_15_1	LayerNorm	1 × 8194 × 1024
63	FFN_15	Feed Forward	1 × 8194 × 1024
64	Attention_16	Multi-Head Attention	1 × 8194 × 1024
65	Add_16	Add	1 × 8194 × 1024
66	LayerNorm_16_1	LayerNorm	1 × 8194 × 1024
67	FFN_16	Feed Forward	1 × 8194 × 1024
68	Attention_17	Multi-Head Attention	1 × 8194 × 1024
69	Add_17	Add	1 × 8194 × 1024
70	LayerNorm_17_1	LayerNorm	1 × 8194 × 1024
71	FFN_17	Feed Forward	1 × 8194 × 1024
72	Attention_18	Multi-Head Attention	1 × 8194 × 1024
73	Add_18	Add	1 × 8194 × 1024
74	LayerNorm_18_1	LayerNorm	1 × 8194 × 1024
75	FFN_18	Feed Forward	1 × 8194 × 1024
76	Attention_19	Multi-Head Attention	1 × 8194 × 1024
77	Add_19	Add	1 × 8194 × 1024
78	LayerNorm_19_1	LayerNorm	1 × 8194 × 1024
79	FFN_19	Feed Forward	1 × 8194 × 1024
80	Attention_20	Multi-Head Attention	1 × 8194 × 1024
81	Add_20	Add	1 × 8194 × 1024
82	LayerNorm_20_1	LayerNorm	1 × 8194 × 1024
83	FFN_20	Feed Forward	1 × 8194 × 1024
84	Attention_21	Multi-Head Attention	1 × 8194 × 1024
85	Add_21	Add	1 × 8194 × 1024
86	LayerNorm_21_1	LayerNorm	1 × 8194 × 1024
87	FFN_21	Feed Forward	1 × 8194 × 1024
88	Attention_22	Multi-Head Attention	1 × 8194 × 1024
89	Add_22	Add	1 × 8194 × 1024
90	LayerNorm_22_1	LayerNorm	1 × 8194 × 1024
91	FFN_22	Feed Forward	1 × 8194 × 1024
92	Attention_23	Multi-Head Attention	1 × 8194 × 1024
93	Add_23	Add	1 × 8194 × 1024
94	LayerNorm_23_1	LayerNorm	1 × 8194 × 1024
95	FFN_23	Feed Forward	1 × 8194 × 1024
96	Attention_24	Multi-Head Attention	1 × 8194 × 1024
97	Add_24	Add	1 × 8194 × 1024
98	LayerNorm_24_1	LayerNorm	1 × 8194 × 1024
99	FFN_24	Feed Forward	1 × 8194 × 1024
100	Output	Output	1 × 8194 × 1024

What the verifier says

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 jinaai/jina-embeddings-v3 --plan --share

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