DINUM / beta.gouv.fr (French Government)

license

Apache-2.0(OSI

source

public

core features withheld

no — publicly funded, README documents full Docker self-hosting

Compar:IA carries a permissive OSI license, Apache-2.0, and the full source is public. The README describes it as 'the open-source LLM arena,' documents Docker self-hosting for the whole platform, and offers no paid tier to withhold anything from that release.

• https://github.com/betagouv/ComparIA recorded 2026-06-22

  Apache-2.0 license badge, full source, betagouv ownership

• https://raw.githubusercontent.com/betagouv/ComparIA/develop/LICENSE recorded 2026-06-22

  LICENSE body 'Apache License, Version 2.0'

• https://raw.githubusercontent.com/betagouv/ComparIA/develop/LICENSE recorded 2026-09-22

  LICENSE body is 'Apache License, Version 2.0'.

• https://raw.githubusercontent.com/betagouv/ComparIA/develop/README.md recorded 2026-09-22

  README states 'Self-host with Docker (single server, automatic HTTPS via Caddy)' and links devops/standalone_docker_install/DOCKER_INSTALL.md for the whole application. It also states 'compar:IA is funded by DINUM and the French Ministry of Culture, with European support from ALT-EDIC' and asks for funders - there is no commercial tier to withhold anything for.

• https://raw.githubusercontent.com/betagouv/ComparIA/develop/README.md recorded 2026-09-22

  README leads 'The open-source LLM arena. Deploy it for your organisation, sector, or language', documents 'Self-host with Docker (single server, automatic HTTPS via Caddy)' for the whole platform, and says 'The whole platform is open source'. No commercial tier, license key or enterprise edition appears in it.

compar:IA's GitHub star count reflects only the repository, not the hosted arena where people actually use the product, so it does not measure reach on its own. The reading instead follows the arena's own collected data: the free-form prompts and preference votes recorded in the team's arXiv paper as of 2026-02-07, most of it in French, rather than a repository or user count.

• https://arxiv.org/abs/2602.06669 recorded 2026-10-08

  600,000+ free-form prompts and 250,000+ preference votes collected as of 2026-02-07, ~89% French

• https://github.com/betagouv/ComparIA recorded 2026-06-22

  74 stars, 16 forks

Compar:IA covers the core arena workflow LMArena offers, blind pairwise voting and a public leaderboard, and adds per-conversation environmental estimates, open datasets published to Hugging Face, data.gouv.fr and Mozilla, and Docker self-hosting. It compares large language models only, without LMArena's image and code arenas.

• https://arxiv.org/abs/2602.06669 recorded 2026-09-29

  Abstract describes "a blind pairwise comparison interface to capture unconstrained, real-world prompts" and releases "three complementary datasets" with "a French-language model leaderboard".

• https://comparia.beta.gouv.fr/ recorded 2026-09-29

  Homepage says users chat with two anonymous AIs, and offers "Mesurer l’empreinte écologique des questions posées aux IA" plus a ranking explainer "Comment le classement compar:IA est-il calculé ?".

• https://github.com/betagouv/ComparIA recorded 2026-06-22

  Feature list: model comparison, voting, EcoLogits, dataset publishing, self-hosting

• https://raw.githubusercontent.com/betagouv/ComparIA/develop/README.md recorded 2026-09-29

  README describes a blind arena: "compar:IA is a blind arena for large language models. You type a prompt, two anonymous models reply, and you vote for the answer you prefer". Roadmap Done items include "Dataset publishing pipeline v1", "Leaderboard v1" and "Language / platform-specific model support"; it offers "Self-host with Docker" with "automatic HTTPS via Caddy" and lists a second instance, "AI-arenaen", in Denmark.