The LMSYS Chatbot Arena is a community-driven evaluation platform where users compare language models through pairwise blind comparisons. With over 5 million user votes, the Arena uses an Elo rating system (similar to chess rankings) to rank LLMs based on real-world conversational performance. As of October 2025, the leaderboard tracks 256 models from a variety of organisations, primarily American and Chinese firms and research institutions.

Currently, no Russian-developed LLMs have appeared on the Chatbot Arena leaderboard, despite significant domestic efforts such as GigaChat (developed by Sberbank) and YandexGPT (developed by Yandex). These models have either not been submitted by their developers or have not met the threshold for inclusion on the platform.

Whether a Russian firm can overcome various challenges—including compute constraints due to semiconductor sanctions and talent bottlenecks from brain drain—to produce a model competitive enough to rank in the top 100 provides an interesting signal on the strength of Russia's AI ecosystem. This captures both a technical dimension (capability for effective large-scale model training) and an institutional dimension (whether Russian firms prioritize international competitiveness versus focusing primarily on domestic markets).

Resolution criteria:

This question resolves Yes if, at any point before January 1, 2027, 00:00 UTC, a Russian-developed model appears in the top 100 on the LMSYS Chatbot Arena Overall leaderboard.

Definition of "Russian-developed model":

A model qualifies if:

1. Entity requirement - Developed by an organization that is:

   • ≥50% Russian-owned (by ownership share), OR

   • Headquartered in Russia, OR

   • A Russian government research institution or state-owned enterprise

2. Training requirement - The model was trained primarily from scratch by the Russian entity (not a fine-tune of a foreign base model). "Trained from scratch" means the pre-training phase was conducted by the Russian entity, though the model may use standard publicly available architectures (Transformer, MoE, etc.).

The "Overall" leaderboard refers to the main text-based Arena Elo leaderboard at https://lmarena.ai/ or https://lmsys.org/leaderboard (or successor URLs managed by LMSYS Org) - NOT category-specific leaderboards (coding, vision, hard prompts, etc.) and NOT the separate Russian-focused "LLM Arena" platform

Rankings determined by publicly displayed Elo score and rank position on the leaderboard

A model counts even if it later drops below top 100, as long as it reached top 100 at any point before the deadline

Model identity must be verifiable through: official model cards, technical papers, company announcements, Arena metadata, or credible reporting from established AI research media

Excluded: Fine-tunes of foreign base models (e.g., Russian fine-tune of Llama, Qwen, Mistral) do NOT count

Excluded: Joint ventures with non-Russian entities

Excluded: Acquisitions where a Russian entity buys a company that developed a model already in top 100

Excluded: Anonymous or pseudonymous models unless their Russian origin is publicly revealed before the deadline

Excluded: Models lacking verifiable documentation of Russian development