Analysis of Qwen3.8-27B Performance and Architecture

Analysis of the Vision Tower architecture and benchmark performance of the Qwen3.8-27B model revealed that the model's performance shows distinct differences depending on the data density of the benchmarks.

Vision Tower Architecture Details:

• Number of Layers: 27
• Hidden Dimension: 1152
• Head Configuration: 16 (head_dim 72)
• MLP: 4304
• Projection: 5120 (direct projection without intermediate adapter)
• Parameters: Approximately 421M (accounting for about 1.56% of the total model)

Benchmark Performance Characteristics: The model leads in 'thin' fields with less data, but recorded mid-tier performance in 'dense' fields with many verified models.

• Top Tier (Thin fields): NL2Repo-Bench, QwenSWEBench, CoWorkBench, IFBench, Agent's Last Exam, etc.
• Mid Tier (Dense fields):
  • Terminal-Bench 2.1: 18th out of 25 models (73.0)
  • GPQA Diamond: 17th out of 50 models (89.2)
  • SWE-Bench Pro: 8th out of 20 models (61.7)

Notably, on SWE-Bench Pro, the 27B model scored 61.7 points, demonstrating efficiency reaching approximately 91% of the 2.4T flagship model's performance (67.7 points). This signifies very high performance efficiency relative to the number of parameters.