Qwen3.8-27B-DSpark

Reconstructed from its own config.json with no weights read. 316K 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

4.14B

4,142,179,840 parameters

In the published checkpoint

1.86B

1,857,358,337 scalars · safetensors.total, read 2026-08-29

Delta

+123%

custom-code This repository ships its own modeling code (`auto_map`, e.g. `dspark.py`), so `config.json` names a class in the repo rather than an architecture `transformers` defines. The graph below is what those config keys mean under `transformers` semantics, which is not necessarily what the repo's own file builds. A gap here is a statement about what we read, not about the checkpoint.

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

$80838.69

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

Structure

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

	Layer	Type	Output shape
1	Input	Input	1 × 262144
2	Embedding	Embedding	1 × 262144 × 5120
3	RoPE	RoPE	1 × 262144 × 5120
4	RMSNorm_1_1	RMSNorm	1 × 262144 × 5120
5	Attention_1	Grouped Query Attn	1 × 262144 × 5120
6	Add_1_attn	Add	1 × 262144 × 5120
7	RMSNorm_1_2	RMSNorm	1 × 262144 × 5120
8	FFN_1	SwiGLU	1 × 262144 × 5120
9	Add_1_ffn	Add	1 × 262144 × 5120
10	RMSNorm_2_1	RMSNorm	1 × 262144 × 5120
11	Attention_2	Grouped Query Attn	1 × 262144 × 5120
12	Add_2_attn	Add	1 × 262144 × 5120
13	RMSNorm_2_2	RMSNorm	1 × 262144 × 5120
14	FFN_2	SwiGLU	1 × 262144 × 5120
15	Add_2_ffn	Add	1 × 262144 × 5120
16	RMSNorm_3_1	RMSNorm	1 × 262144 × 5120
17	Attention_3	Grouped Query Attn	1 × 262144 × 5120
18	Add_3_attn	Add	1 × 262144 × 5120
19	RMSNorm_3_2	RMSNorm	1 × 262144 × 5120
20	FFN_3	SwiGLU	1 × 262144 × 5120
21	Add_3_ffn	Add	1 × 262144 × 5120
22	RMSNorm_4_1	RMSNorm	1 × 262144 × 5120
23	Attention_4	Grouped Query Attn	1 × 262144 × 5120
24	Add_4_attn	Add	1 × 262144 × 5120
25	RMSNorm_4_2	RMSNorm	1 × 262144 × 5120
26	FFN_4	SwiGLU	1 × 262144 × 5120
27	Add_4_ffn	Add	1 × 262144 × 5120
28	RMSNorm_5_1	RMSNorm	1 × 262144 × 5120
29	Attention_5	Grouped Query Attn	1 × 262144 × 5120
30	Add_5_attn	Add	1 × 262144 × 5120
31	RMSNorm_5_2	RMSNorm	1 × 262144 × 5120
32	FFN_5	SwiGLU	1 × 262144 × 5120
33	Add_5_ffn	Add	1 × 262144 × 5120
34	Final_RMSNorm	RMSNorm	1 × 262144 × 5120
35	LM_Head	Linear	1 × 262144 × 248320
36	Output	Output	1 × 262144 × 248320

What the verifier says

warnLinear "LM_Head" is 5120 × 248320 = 1271M parameters (~4.7 GB float32). A single dense layer this large usually means a feature map was flattened without pooling first; embedding / vocab-projection heads are the expected exception. Fix: Add a Global Average Pool or more downsampling before the Linear, or factorize it (low-rank / bottleneck projection).
huge-linear-params

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 RadixArk/Qwen3.8-27B-DSpark --plan --share

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

Other qwen3 checkpoints

jina-reranker-v3

596M derived · -0.13% against the checkpoint

Kimi-K3-DSpark

4.26B derived · +89.3% against the checkpoint

Qwen3-0.6B

596M derived · -20.7% against the checkpoint

Qwen3-1.7B

1.72B derived · -15.3% against the checkpoint