AI Potluck
Model components / Evaluation code

VLMEvalKit

OpenCompass Community

Evaluation toolkit for large vision-language models that runs 220+ models over 80+ multimodal benchmarks with one command and generation-based scoring. It produces the numbers behind the OpenVLM Leaderboard, the reference leaderboard for open multimodal models, and is installed from source rather than a package registry.

Kept as a separate product from opencompass: separate repo, separate scope and its own OpenVLM Leaderboard, integrated by OpenCompass as a backend. No PyPI package exists (vlmeval is the import name only; 404 on the JSON API and the simple index). Verified 2026-09-01 via the open-compass/VLMEvalKit repository, its license endpoint, its README and Quickstart.

Openness

5 high confidence
5.0
license
Apache-2.0(OSI)
source
public(GitHub)
core-gated
ungated

Apache-2.0 and fully public. The LICENSE body is the stock Apache-2.0 text under "Copyright 2023 VLMEvalKit Authors" with no appended terms. Install is git clone + `pip install -e .`; the README documents no pricing, paid tier, enterprise edition or license key — the only commercial references are the commercial model APIs it can evaluate.

Adoption

3 medium confidence
3.0

The toolkit behind the OpenVLM Leaderboard and OpenVLM Video Leaderboard (Hugging Face spaces under the opencompass org), with 4,370 stars and results published for 220+ models — the reference evaluation surface for open multimodal models, and integrated as a backend by both OpenCompass and EvalScope. No download channel exists to count: there is no package (vlmeval is the import name only, PyPI 404 this session) and the install path is a git clone, so the level rests on reported traction rather than stars alone, the same shape as opencompass's own record. Conservative alternative if the writer prefers the mechanical instrument: stars_fallback at level 2 (1K-10K stars).

Capability

4 medium confidence
4.0

The multimodal peer of its own org's flagship: comparable breadth (220+ models, 80+ benchmarks against opencompass's 70+ datasets), the same leaderboard-producing role (OpenVLM against CompassRank), scoped to vision-language models rather than the cross-domain standardization reference — so at opencompass's 4, one below lm-evaluation-harness.

Verified 2026-09-01