howtospark
models / qwen3-5-35b-a3b

Qwen3.5 35B-A3B

Qwen/Qwen3.5-35B-A3B
params
36.0B
active params
3.0B
sparsity
MoE
attention
Full attention
architecture
qwen3_5_moe
license
apache-2.0
downloads
2.3M
likes
1.5k
hf updated
Apr 24, 2026

Visualize

total size

each column is a layer — lit blocks are the experts that fire per token

bits / param
16-bit
footprint
67.0 GiB
sparks to fit
block ≈
281M params
active paramsdormant expertsghost = precision traded awayblock fill ∝ bits kept · Qwen3.5 35B-A3B35.951822704B params

Eval scores (compare all)

Reported

85.3MMLU-Pro
84.2GPQA Diamond
74.6LiveCodeBench
69.2SWE-bench Verified