EVA¶

Development line: project:eva · thread eva-public-resources
Last event: 2026-03-26 · 1 dated since 2026-03-26 · Researched: 2026-09-05 · confidence: medium

What it is¶

EVA (Efficient Video Agent) is a video agent built on Qwen2.5-VL for researchers and engineers.

• Agent cycle: runs a summary → plan → action → reflection loop.
• Frame selection: picks and extracts video frames, then updates the next step from the result.
• Model weights: hosted at WRHC/EfficientVideoAgent for evaluation across six video datasets.

The published repository covers multi-GPU evaluation through vLLM rather than an interactive interface. The setup works for reproducible query-driven video evaluation, but practical serving requires a separate inference pipeline.

Development line¶

• 2026-03-26 — EVA source and model resources were recorded. On 2026-03-26, EVA appeared with a GitHub source repository and a Hugging Face model page. Both links mark public access for the project. The dated links alone do not confirm a version, capabilities, evaluation results, or whether either resource was first published that day.

What changed¶

2026-03-26 — Official evaluation code and EVA weights are public. The model replaces passive uniform frame sampling with an agent loop that chooses what to inspect and when.

How to use this¶

As of 2026-03-26, practitioners should treat EVA as having both a source-code reference and a model-hosting reference, then verify setup and model details directly from those resources before use.

1. Download weights for WRHC/EfficientVideoAgent and install the repository dependencies with FFmpeg. — https://github.com/wangruohui/EfficientVideoAgent
2. Prepare one of the six supported video datasets for evaluation and set the local video_root in DATASET_CONFIG. — https://github.com/wangruohui/EfficientVideoAgent
3. Serve the weights with a vLLM OpenAI-compatible endpoint, then set its URL, tokenizer path, and allowed media paths in eval-eva.py. — https://huggingface.co/WRHC/EfficientVideoAgent
4. Run the chosen dataset with eval-eva.py. Re-run the same command to resume an interrupted run from the cache file. — https://github.com/wangruohui/EfficientVideoAgent

Best practices¶

• Keep ffprobe in PATH and FFmpeg libraries in LD_LIBRARY_PATH because frame extraction runs during evaluation. — https://github.com/wangruohui/EfficientVideoAgent
• Grant vLLM access to local video files with --allowed-local-media-path so the agent can pass extracted frames to the model. — https://huggingface.co/WRHC/EfficientVideoAgent
• Lock --max-visual-tokens across runs because exceeding the token budget lowers resolution and drops frames. — https://huggingface.co/WRHC/EfficientVideoAgent
• Do not treat a single run as an exact reproduction because accuracy varies by fractions of a percent even at temperature=0. — https://github.com/wangruohui/EfficientVideoAgent

Superseded by this¶

• Nothing marked obsolete yet.

Still unknown¶

• No dated first-party changelog exists after the initial release. We cannot confirm whether later commits or model updates changed EVA's capabilities.
• Public documentation covers only the benchmark evaluation workflow. We find no confirmed production inference setup, standalone API, or user interface.

Sources¶

Agent brief¶

• Subject: project:eva, thread eva-public-resources, 1 dated events 2026-03-26 → 2026-03-26.
• Practical note: As of 2026-03-26, practitioners should treat EVA as having both a source-code reference and a model-hosting reference, then verify setup and model details directly from those resources before use.
• Confidence: medium. Dated supersedes above are the authority for what is obsolete.