Dataset Viewer
Auto-converted to Parquet Duplicate
model
stringclasses
34 values
model_slug
stringclasses
34 values
revision
stringclasses
34 values
tensor
stringclasses
4 values
layer_a
int64
0
92
layer_b
int64
1
93
offset
int64
1
93
cosine
float64
-0.05
0.99
row_mean_cosine
float64
-0.07
0.98
⌀
both_inside_region
bool
2 classes
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
1
1
-0.000845
-0.000757
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
2
2
-0.000222
-0.000264
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
3
3
-0.000255
-0.000241
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
4
4
-0.000085
-0.000056
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
5
5
0.000315
0.000369
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
6
6
-0.0001
-0.00008
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
7
7
0.000181
-0.000071
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
8
8
-0.000176
-0.000318
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
9
9
0.000004
-0.000073
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
10
10
-0.000413
-0.000522
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
11
11
-0.000372
-0.000434
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
12
12
-0.000536
-0.000446
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
13
13
-0.000014
0.000022
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
14
14
-0.000211
-0.000194
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
15
15
0.000223
0.000262
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
16
16
-0.00076
-0.000688
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
17
17
0.000585
0.00069
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
18
18
-0.000013
-0.000008
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
19
19
-0.000989
-0.000976
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
20
20
0.000568
0.00068
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
21
21
-0.000115
-0.000017
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
22
22
0.000478
0.000332
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
23
23
0.000113
0.000103
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
24
24
-0.000591
-0.000712
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
25
25
-0.000855
-0.000801
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
26
26
0.000004
-0.000031
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
27
27
-0.000245
-0.000207
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
28
28
0.000442
0.000353
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
29
29
-0.000619
-0.000546
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
30
30
0.000172
0.000244
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
31
31
0.00035
0.000384
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
32
32
-0.000501
-0.000512
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
33
33
0.000159
0.000069
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
34
34
0.000121
0.000202
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
35
35
0.000047
0.000128
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
36
36
-0.000532
-0.000313
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
37
37
0.000368
0.00034
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
38
38
-0.000199
-0.000176
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
39
39
-0.000068
-0.000208
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
40
40
-0.000424
-0.000428
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
41
41
-0.00041
-0.000476
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
42
42
0.00002
0.000097
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
43
43
-0.000939
-0.00114
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
44
44
-0.000616
-0.000095
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
45
45
0.000651
0.000588
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
46
46
0.000597
0.000621
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
47
47
0.000342
0.000399
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
48
48
0.000806
0.000796
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
49
49
-0.000827
-0.000727
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
50
50
0.000059
-0.000098
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
51
51
-0.000489
-0.000365
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
52
52
-0.00023
-0.000471
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
53
53
0.000414
0.000407
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
54
54
-0.000431
-0.000395
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
55
55
0.000246
0.000304
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
56
56
-0.000146
-0.000102
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
57
57
-0.000256
-0.000198
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
58
58
-0.00029
-0.000241
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
59
59
0.000774
0.000963
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
60
60
-0.000044
-0.00008
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
61
61
-0.000169
-0.000102
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
62
62
0.000074
0.000244
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
0
63
63
0.000513
0.000549
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
2
1
-0.000589
-0.00064
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
3
2
-0.000631
-0.00073
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
4
3
0.000235
0.000229
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
5
4
0.000131
0.000123
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
6
5
0.000846
0.001028
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
7
6
-0.001221
-0.001297
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
8
7
0.00104
0.001133
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
9
8
-0.000247
-0.000177
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
10
9
0.001237
0.001342
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
11
10
-0.000512
-0.00044
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
12
11
0.000238
0.000275
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
13
12
0.000273
0.000152
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
14
13
-0.000064
0.000063
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
15
14
0.000815
0.0008
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
16
15
0.00069
0.000782
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
17
16
-0.000832
-0.000687
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
18
17
0.000839
0.000709
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
19
18
0.000017
-0.000097
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
20
19
-0.000858
-0.0009
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
21
20
0.000763
0.000988
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
22
21
-0.000167
-0.000194
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
23
22
-0.000368
-0.000371
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
24
23
-0.00082
-0.000908
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
25
24
-0.000063
-0.000013
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
26
25
-0.000214
-0.000304
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
27
26
0.000849
0.00096
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
28
27
-0.000358
-0.000326
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
29
28
0.000086
-0.000033
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
30
29
0.000103
0.000264
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
31
30
-0.000146
-0.00008
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
32
31
0.001068
0.000872
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
33
32
0.000017
-0.00002
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
34
33
0.000458
0.000544
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
35
34
0.000326
0.000327
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
36
35
-0.000311
-0.000285
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
37
36
0.001218
0.001055
false
Qwen/Qwen3-32B
qwen3-32b
9216db5781bf21249d130ec9da846c4624c16137
self_attn.k_proj.weight
1
38
37
0.00103
0.000937
false
End of preview. Expand in Data Studio

Repeated blocks in Qwen-family checkpoints: pinned weights and neuron responses

v0.2.1 packaging note: checksum metadata corrected; measurements unchanged.

New in v0.3.0: Base-reference checks

Using the published Qwen3-14B-Base instead of Qwen3-14B as the reference against the same Qwen3-32B target changes the primary key/value and sampled gate-row mean similarities by less than 0.0002. The best matching layer is unchanged for every source layer in both full key/value maps. A separate scan inside the published Qwen3-30B-A3B-Base finds its strongest eligible offset at 12 layers in both keys and routers. Results, methods and data.

This is a separate weight study. It retains the recorded prediction without scoring it and adds no neuron-response measurements. The earlier studies and original survey measurements are unchanged. v0.3.0 changes.

New in v0.2.0, part 1: the 14B/32B weight maps

Qwen2.5-32B and Qwen3-32B each show an ordered layer-by-layer match to their 14B counterpart, with a middle block matching at two consecutive positions in the 32B. These maps are consistent with expansion from 14B-sized models by block duplication, a construction method disclosed for SOLAR and KAT. The layer maps are measured directly between pinned checkpoints (fixed commits of the model repositories), with additional neuron/head measurements for Qwen3.

Family Decoder layers, 14B → 32B Measured repeated range inside the 32B MLP width, 14B → 32B Findings
Qwen2.5 48 → 64 16–31 ↔ 32–47 13,824 → 27,648 Cross-checkpoint keys, biases and norms
Qwen3 40 → 64 8–31 ↔ 32–55 17,408 → 25,600 MLP weights, cross-checkpoint gate rows and head maps

For Qwen2.5, the 14B's layers 16–31 match both consecutive 32B blocks: mean key-weight cosine 0.485 and 0.738, and query-bias cosine 0.993 and 0.998. The maximum absolute key cosine off that map is 0.004315. The 14B's layers 0–15 match 32B layers 0–15, and its layers 32–47 match 32B layers 48–63.

For Qwen3, the first 17,408 gate-neuron indices show substantial correspondence with the 14B's, and the 8,192 higher-index channels have a structured relationship to leading channels. Early-layer exceptions and sampled head-map exceptions are retained in the results.

Two measurements are reported for Qwen3. Inside Qwen3-32B, layers 32–55 match layers 8–31. Between checkpoints, 14B layers 0–9 match 32B layers 0–9 best, while 14B layers 10–39 match 32B layers 34–63 best (an offset of 24). For 14B layers 8 and 9 the first-position block is closer; from layer 10 onward the second-position block is closer. These comparisons describe different maps.

Expansion is the interpretation of these measured maps. The released checkpoints do not determine which checkpoint came first or when a construction step occurred. The Qwen2.5 MLP width is known to double; its channel-by-channel widening procedure has not yet been measured.

See the short glossary for the measurement terms used here.

New in v0.2.0, part 2: do the corresponding neurons respond alike?

This is a separate measurement on a different artifact: activations of the official Qwen3-32B Q8 GGUF checkpoint on a fixed text corpus, not the stored BF16 weights compared above. It asks only how alike same-index neurons 24 layers apart respond. It makes no claim about the weight maps and does not establish a shared function.

In that study (one Apple M5 Max, 48 GB, llama.cpp with Metal; about two hours), same-index partners rank first of 101 candidates in 185/288 sampled pairs. The typical resemblance is small: median final-activation correlation 0.0555, median overlap 4 of the top 100 windows and 0 of the top 10. Response results and full evidence.

Where to start

Study Included evidence Start here
Qwen2.5-14B ↔ 32B weights Six 48×64 cosine matrices, all 18,432 pair values, original capture, pinned tensor hashes and CPU reproduction Qwen2.5 findings
Qwen3-32B MLP and Qwen3-14B ↔ 32B weights All-pair MLP measurements, K/V maps, aligned gate rows, higher-index-channel and head checks Qwen3 findings
Qwen3-32B Q8 responses All 288 sampled pairs, 10,000 fixed windows, full retained measurements and 512 raw paired-channel windows Response findings
Qwen3 Base-reference follow-up Six full K/V maps, 544 sampled gate rows, Base/non-Base calibration and Base 30B within-checkpoint scan Base follow-up
Original 34-checkpoint survey Original within-checkpoint K/V, router and convolution tables, scanner and figures Survey results

Read the tables and plots online without downloading the large arrays. The weight studies' offline table checks need under 70 MB of study files, plus Python dependencies. Fresh tensor remeasurement is a separate download. Reading the response tables is also lightweight, but re-deriving them needs the retained capture files, about 3.5 GB (3.3 GiB). No model weights are bundled. It uses the official Q8 GGUF checkpoint (Qwen’s published 8-bit quantization), separately pinned from the BF16 weight scans. The corpus has 10,000 distinct windows; an article-title parser defect is documented separately.

Changes and scope · Reproduction

Original weight survey

The following survey and its root data/ tables retain the v0.1.2 measurements. The new studies are stored separately under studies/; the survey counts are not expanded by treating cross-checkpoint pairs or activation samples as additional checkpoints.

Several Qwen checkpoints contain a stretch of decoder blocks whose weights are far more similar, block for block, to the stretch a fixed offset away than any other pair of blocks at that offset: layers 32–55 of Qwen3-32B line up with layers 8–31, layers 20–31 of Qwen3.5/3.6-35B-A3B line up with 8–19, layers 32–47 of Qwen2.5-32B line up with 16–31, and so on across the surveyed checkpoints. Corresponding pairs show elevated cosine relative to the other pairs at the same offset. Routers also have a nonzero background, and some checkpoints have additional structure at other offsets. SOLAR-10.7B provides a disclosed depth-upscaling comparison in the same tensor family. This release measures that pattern in 34 pinned checkpoints with nothing but HTTP byte-range reads of their public weight files, and ships the data, the scanner and a verifier so anyone can check it on a laptop. For survey counting, grouping related variants and derivatives gives 16 Qwen or Qwen-derived checkpoints with a block (about 9 groups of related checkpoints) and 17 without (16 grouped rows), plus SOLAR as a separate calibration. The grouped positive rows are Qwen3.5-35B Base/Instruct with Qwen3.6-35B, Qwen3.5/3.6/3.8-27B, Qwen3-30B with its 2507 variant, and Qwen2.5-32B with R1 and KAT; the negative rows group Qwen2.5-7B with Math-7B. Every checkpoint is measured separately.

Jeffrey Shorthill · independent interpretability research · 2026-10-27 · v0.3.0

Repeated blocks across the Qwen family

Six survey rows at a glance

Within-block mean cosine between corresponding layers, with the mean over the remaining pairs at the same offset for comparison. Both are flattened float64 cosines except the explicitly labeled router-row values, which use aligned per-row cosines for both means. Full table: RESULTS.md.

Checkpoint Layers Corresponding blocks Tensor Within-block Same-offset, off-block
Qwen3-32B 64 8–31 ↔ 32–55 attention k_proj / v_proj 0.254 (n=24) / 0.447 (n=24) 0.000 (n=16) / 0.000 (n=16)
Qwen3.6-35B-A3B (Qwen3.5-35B-A3B and its Base within 0.002) 40 8–19 ↔ 20–31 DeltaNet conv1d / k_proj / router rows 0.896 (n=9) / 0.524 (n=3) / 0.795 (n=12) -0.001 (n=12) / 0.000 (n=4) / 0.260 (n=16)
Qwen2.5-32B (and DeepSeek-R1-Distill-Qwen-32B) 64 16–31 ↔ 32–47 attention k_proj; v_proj (2.5-32B only) 0.424 / 0.422 (n=16 each); v_proj 0.544 (n=16) 0.000 (n=32 each); v_proj 0.000 (n=32)
Qwen3.6-27B (3.5/3.8 measured separately; values here are 3.6) 64 8–23 ↔ 24–39 ↔ 40–55 DeltaNet conv1d / k_proj 0.859 (n=24) / 0.309 (n=8) 0.001 (n=12) / 0.000 (n=4)
SOLAR-10.7B-v1.0 (disclosed up-scaling, calibration) 48 8–23 ↔ 24–39 attention k_proj 0.871 (n=16) 0.000 (n=16)
KAT-V1-40B (Qwen2.5-32B expanded, per its report) 80 24–31 ↔ 32–39 and 40–47 ↔ 48–55 attention k_proj 0.984 / 0.978 (n=8 each; 16 total) 0.000 (n=56)

Counts refer to the displayed means. In the full verdict table, “confirmed” means passing the reporting rule for the declared pattern, not a demonstrated training history. The 27B row pools the two consecutive block comparisons: 12 convolution pairs and 4 attention pairs per comparison. KAT displays its two separate 8-pair means separately; Qwen2.5-32B, its R1 derivative, and SOLAR contribute 16 pairs per 16-layer segment comparison. The value projection (v_proj) was added for Qwen3-32B and Qwen2.5-32B on 2026-10-21 as a second attention tensor family. Measured against its own same-offset background, it shows elevated similarity at the same corresponding positions as the key projection (within-block 0.447 and 0.544, off-block 0.000, offset spectrum peaking only at K, the declared block spacing), and in both checkpoints, corresponding value projections have higher within-block cosine similarity than corresponding key projections. Keys and values are trained together, so this corroborates rather than independently replicates the key-projection result; the query and output projections and the MLP are outside this original survey table. The v0.2.0 Qwen3-32B MLP and cross-checkpoint measurements are reported in the separate weight study linked above.

Under the inherited-block interpretation, KAT also predicts offsets 16+8=24 and 16+2×8=32; the four observed 8-pair correspondences are listed in RESULTS.md.

Across the surveyed MoEs, flattened router means at offset 1 round to 0.26–0.51. The table and chart compare within-block similarity against the same-offset off-block background; the raw-rank reporting rule's margins are reported in RESULTS.md. The checkpoints with no repeat (every model at or below 14B, Qwen3.5-122B and 397B, Qwen3-Coder-480B, Qwen1.5/2 MoEs) show no eligible neighboring-offset spike above 0.010175 under the reporting rules. Two small-hybrid attention families have no offset with six pairs and remain explicitly unranked; their convolution families support the model-level no-block result.

Check it yourself

The full-package check below expects the complete snapshot, including the response archive. For small weight-only downloads and reductions, use the selective route.

Offline, allow a few minutes on an SSD, no network: verifies every file hash, the lock and scanner hashes, the block coverage against each pinned config.json, and that RESULTS.md regenerates byte for byte.

git lfs install
git clone https://huggingface.co/datasets/ec75hash/copy-blocks-qwen
cd copy-blocks-qwen
python3 -m venv .venv && source .venv/bin/activate && pip install -r requirements.txt
python -m unittest discover -s tests
python scripts/verify_release.py
python scripts/make_report.py --out /tmp/figs --report /tmp/RESULTS.md

Online, the real test: re-read the pinned tensors from the Hub and recompute everything. The scanner downloads only config.json, the safetensors index (or header) and the selected tensors' byte ranges; it refuses any response that is not an honoured Range request. No model runs. No GPU. This example selects Qwen3-32B, Qwen3.6-35B-A3B, Qwen2.5-32B and the SOLAR calibration (about 3.2 GB of tensor payloads, including k_proj and v_proj for the two dense checkpoints):

python scripts/scan_weights.py --models qwen3-32b qwen36-35b-a3b qwen25-32b solar-10p7b \
    --out reproduction/data --provenance reproduction/metadata
python scripts/make_report.py --data reproduction/data --out reproduction/figures --report reproduction/RESULTS.md
python - <<'PY'
import csv, math
selected = {"Qwen/Qwen3-32B", "Qwen/Qwen3.6-35B-A3B",
            "Qwen/Qwen2.5-32B", "upstage/SOLAR-10.7B-v1.0"}
def read(path):
    with open(path) as f:
        rows = [r for r in csv.DictReader(f) if r["model"] in selected]
    pairs = {(r["model"], r["revision"], r["tensor"], r["layer_a"], r["layer_b"]):
             float(r["cosine"]) for r in rows}
    assert len(pairs) == len(rows), "duplicate pairs"
    return pairs
a, b = read("data/pairs.csv"), read("reproduction/data/pairs.csv")
assert a.keys() == b.keys(), "missing or extra pairs"
assert all(math.isfinite(b[k]) and abs(a[k] - b[k]) <= 1e-9 for k in a)
print("Matched", len(a), "pairs within 1e-9")
PY

The original survey covers 34 checkpoints and 2,464 tensors (11.1 GB of tensor payloads). The original v0.1.0 capture covered 2,336 tensors (9.8 GB); its recompute from the local cache took 13 minutes. The v0.1.1 extension adds 128 value-projection tensors (1.34 GB). Hardware used here: Apple M5 Max, 48 GB, macOS 27.0.1, Python 3.14.4, NumPy 2.4.2 (Accelerate BLAS). The largest normalized input stack is 80 × 5,242,880 values in float64 (3.36 GB); peak memory is higher because decoded weights, stacked copies and conversion temporaries coexist. The complete survey was run with 48 GB of RAM.

Independent reproduction

A second scanner, written separately in a Codex (GPT-6) session with its own tensor cache and its own pair table, measured 33 of these checkpoints at the same revisions; its roster omitted DeepSeek-R1-Distill-Qwen-32B. All 38,540 layer pairs the two tables share agree to within 4 × 10⁻¹⁵, and all 1,552 shared tensor payload hashes match. Details in docs/INDEPENDENT_CHECK.md. That comparison also caught a real bug: on this platform a float32 Gram matrix of 32 or more multi-million-element rows is wrong by about 0.005 (normalised diagonals come out 0.99). Both scanners now accumulate in float64, with diagonal tolerances of 10⁻⁹ here and 10⁻¹⁰ in the independent scan. The largest correction in the earlier Qwen3.6-27B release is SOLAR's within-block mean, 0.870722 → 0.871450; the smaller 27B changes are listed in the independent-check note.

Limitations and notes

The within-block similarity in the table above, measured against the same-offset background and the surveyed controls, is consistent with the paired blocks having started from the same weights, as in the two disclosed depth up-scalings in the table (SOLAR, KAT). Weights alone do not show how or when a model was built, how much training followed any copy, or whether paired blocks now compute the same function. Cosine is not a clock.

  • The original survey is weights only. The Q8 response study adds measured activations under its own scope, without a same-feature or causal-function verdict.
  • Shared initialisation is the reading, not a recipe. Qwen has not published how these checkpoints were built. In the original survey the reading "copied, then trained" rests on the SOLAR calibration and the off-block floor; for Qwen2.5-32B and Qwen3-32B the v0.2.0 cross-checkpoint studies add a trained external reference.
  • One tensor family per claim (two, k_proj and v_proj, for Qwen3-32B and Qwen2.5-32B). Dense models are read through k_proj; hybrids add the DeltaNet conv1d; MoEs add the router. Other tensors may tell a more or less similar story.
  • Routers have a non-zero baseline between any two layers, so their numbers are only meaningful against the same-offset off-block value printed beside them.
  • No confidence intervals: rows within a checkpoint share weights and are not independent samples. The measurement is deterministic; the same bytes give the same numbers.
  • Pinned revisions only (models.lock.json). Qwen3-Next-80B-A3B and Qwen3-Coder-480B-A35B are read from their Instruct repositories because no base checkpoint is published.
  • KAT-V1-40B's Hub index names the wrong shard for 76 of its 80 k_proj tensors; the scanner resolves tensors against the shard headers and records the mismatch in data/provenance.json.

Original survey files and provenance

The Hub dataset is the canonical versioned release. The GitHub repository mirrors code and documentation with the two small weight studies, without the large response arrays.

  • models.lock.json: 34 checkpoints, full commit SHAs, SHA-256 of each config.json and index, tensor families, and the block hypothesis (period and region, or explicit segments) per row. Its note fields are historical working annotations: “chart” refers to the September 2026 survey chart described in chart corrections; the remaining notes describe controls, variants and checkpoint selection. These notes are not measurement inputs.
  • data/pairs.csv: 60,982 layer pairs: model, revision, tensor, layer_a, layer_b, offset, cosine, row-mean cosine, whether the pair is a corresponding-block pair. The historical both_inside_region field marks aligned-pair membership here; column definitions explain this and the other fields.
  • data/offsets.csv, data/summary.json: per-offset means and the per-family summaries. The summary's inside_row_mean and outside_row_mean provide the matched row-mean comparison shown in RESULTS and the README router row (0.795 versus 0.260 for Qwen3.6-35B).
  • data/provenance.json: every tensor's shard, dtype, shape, absolute byte range, payload and header SHA-256, plus scanner and lock hashes, package versions and run time. Capture-time hashes are retained alongside recorded reporting amendments; the added off-block row means were aggregated from the saved pair table without a new weight scan.
  • provenance/metadata/: the pinned config.json and index files (no tensor data).
  • checksums.json: SHA-256 of every released file except the manifest itself; scripts/verify_release.py checks it.
  • figures/: the family chart, offset spectra for the 37 confirmed families, and six heat maps.
  • METHODS.md · RESULTS.md · docs/CHART_CORRECTIONS.md (what changed against the author's earlier hand-assembled chart) · NOTICE · CITATION.cff

Earlier related release: Qwen3.6-27B weight and readout measurements, https://github.com/ec75hash/qwen-depth-upscaling-jpp-check.

Code Apache-2.0 (LICENSE); original derived data, figures and prose CC BY 4.0 (LICENSE-DATA.md). WikiText-derived excerpts retain their upstream declarations and attribution. The scanner derives from the 27B release; Claude assisted with its extensions, the report code and this documentation; the independent check was a Codex session. The survey, its direction and its interpretation are the author's.

Downloads last month
543