Qwen2.5-VL-7B-Instruct vs GLM-5.3-Flash

Try Try

Qwen2.5-VL-7B-Instruct has 8.29B parameters (7.07B active per token) and takes 16.6 GB on disk; GLM-5.3-Flash has 321B (16.7B active) and takes 328 GB. Compare them layer by layer.

At a glance

Qwen/Qwen2.5-VL-7B-Instructzai-org/GLM-5.3-Flash
Repo createdJan 26, 2025updated Apr 6, 2025Aug 25, 2026updated Sep 7, 2026
Model typeDenseFull attention (GQA)Mixture of expertsHybrid attention: 34 linear + 12 full (MLA)
InputsText + images677M vision encoderText + images564M vision encoder
Total parameters8.29B39× less321B39× more
Active per token7.07B2.4× less85% of the model16.7B2.4× more5.2% of the model
ExpertsNone (dense)8 of 288 activeplus 1 shared, always on
Max context (from config)125K tokens8.2× less1M tokens8.2× more
Layers281.6× less451.6× more34 linear + 12 full attention · plus 1 extra prediction layer
On disk16.6 GB20× less5 files328 GB20× more62 files
PrecisionBF16 (100%)FP8 E4M3 (96%), BF16 (4.2%)
QuantizationNoneoriginal precision (BF16)FP8 E4M3blocks of 128×128 · 98% of parameters · attention, embeddings & output head and vision encoder kept in BF16
Fewest GPUs1× RTX 4090a single 24 GB card · weights only5× NVIDIA H100fits in one 8-GPU server · weights only
Licenseapache-2.0mit
GitHubQwenLM/Qwen2.5-VLNot linked