howtospark
models / qwen3-5-27b

Qwen3.5 27B

Qwen/Qwen3.5-27B
params
27.8B
sparsity
Dense
attention
Full attention
architecture
qwen3_5
license
apache-2.0
downloads
2.6M
likes
1.0k
hf updated
Apr 24, 2026

Visualize

total size

each column is a layer — every parameter fires on every token

bits / param
16-bit
footprint
51.7 GiB
sparks to fit
1×
block ≈
217M params
active paramsghost = precision traded awayblock fill ∝ bits kept · Qwen3.5 27B ≈ 27.781427952B params

Eval scores (compare all)

Reported

72.4SWE-bench Verified
86.1MMLU-Pro
85.5GPQA Diamond
80.7LiveCodeBench