We used HFlow to evaluate the latest open weights VLMs for processing egocentric data

View original

GeminiGitHubModel releaseModel accessOpen source

Heat trend

Collecting trend data

The percentage is based on available heat signal, not comment count or independent people.

Why it matters

Gemini model activity is surfacing — worth tracking for capability changes, ecosystem impact, and availability.

AI summary

HFlow was utilized to evaluate open-weight Vision-Language Models (VLMs) for egocentric data processing. This evaluation mirrored Build AI's Egocentric-10k, employing identical prompts and datasets, but varying the VLM. The baseline model, Gemini 2.5 Flash, achieved a 91.65% score in measuring hand visibility and active manipulation. The HFlow tool is available via git clone https://github.com/Hebbian-Robotics/hflow.git.