Does gUrrT have a benchmark to compete against?

Before I answer that, let me explain what gUrrT actually is and what problem it is solving. Largely, Video Understanding has been solved by Video Language Models. What are Video Language Models? They are largely LLMs tuned to handle video context, you pass in a video, its context is extracted, and an LLM with its existing knowledge base does wonders. Open source models like Qwen2.5-VL 7B require ~18 GB at fp16. InternVL3–40B and 72B-class models require 80 GB+, nowhere possible on a consumer grade GPU (~8 to 12 GB). Even with quantization, the length of video these models can handle and the inference time are a real big bottleneck. Then there is Claude, it does not natively accept video files. No video understanding out of the box. Then there is Gemini. Free tier does process video. it actually sees your content. But the video limit is 5 minutes. Your lectures are going onto Google’s data centers. Rate limited on free. And the moment you want anything serious, you are on a paid plan, where your videos are still going to Google. Then comes premium Gemini 3 Pro, supports video up to 1 hour, but it comes at a cost. Everything is running on Google’s infrastructure. You are feeding your videos and lectures onto Google’s data centers. 1 hour duration limit. Still a bottleneck.

All of this just for the LLM to retrieve the right context from video.

So gUrrT plays it smart, it asks a question what if the context can be built and let the LLM keep doing its job? This is the sweet spot where we are doing our job. Smartly extract only the frames where content genuinely changed. Transcribe the audio via STT. Index everything into a searchable vector store. Then hand that context to a state of the art LLM for reasoning no LVLM in the pipeline, no 80 GB GPU, no video leaving your machine. The LLM does not need to understand video. It just needs the right context. That is the problem gUrrT solves.

This brings me to benchmarks. We cannot run state of the art video language models on consumer hardware they simply will not fit. The VRAM wall is real. The models that can actually handle video Gemini, Claude, ChatGPT are behind a paywall. And even then, you are rate limited, duration limited, and your video is leaving your machine. Then there are the small open source models that do fit on consumer VRAM. MiniCPM-V 4.6 runs in 4 GB. SmolVLM2 runs in under 1 GB. They run but they are capped at 50 to 128 frames per video. A 30min lecture at 1 frame per second is 1,800frames. These models see only 7% of it. They are not understanding the video, they are guessing from a sparse sample.

What I can tell you is how gUrrT performs currently how much has it improved from v1 to v2.

A sneak peek at how gUrrT v2 compares to v1.

v2 RELEASING SOON!!!!!!