So... has anyone actually figured out whose model Elephant Alpha is yet?

A powerful and anonymous AI model dubbed 'Elephant Alpha' has surged to the #1 spot on the OpenRouter leaderboard, processing an impressive ~250 tokens per second (tps). The model boasts a massive 100 billion parameters and a 256K token context window, but its performance is highly specialized. It demonstrates notoriously poor Chinese language support while being heavily optimized for coding and 'agentic' tasks—AI that can take actions and follow instructions with remarkable strictness. This unique performance profile, combined with the sheer compute power required to serve it at this speed for free, has the AI community buzzing with speculation about its origins.

The model's architecture and capabilities have effectively ruled out several major players. Its weak Chinese support makes it unlikely to be from Chinese labs like Qwen or DeepSeek. It also doesn't align with Google's typical transparency about model sizes. The shortlist of entities capable of training and serving such a model at this scale points to a few possibilities: a new, highly optimized variant of Cohere's Command model, a sophisticated Mixture of Experts (MoE) system from an emerging, well-funded startup, or a stealth project from an established tech giant. The mystery underscores the increasingly competitive and opaque nature of frontier AI development.