Gemini 3.8 Flash Leads Web Access on tb4.0 Benchmark
Research engineer Florian Brand shared benchmark data on model tool usage.
TLDR
Florian Brand, research engineer at Prime Intellect, posted that Gemini 3.8 Flash uses web access tools such as web search, curl, git clone, and HEAD requests more than all other models combined on the tb4.0 benchmark. He listed GLM-5.3 and Gemini 3.8 Flash as models that perform these actions. The post links to a horizontal bar chart titled "Web access by recorded acting model" with subtitle "Failed and unresolved calls."
Combined views
2.6K
2 Sources, first seen 24d ago