01.AI

Contribution to AI

A 34-billion-parameter model that outscored a 70-billion one made a point worth more than the ranking. The Yi-6B and Yi-34B base models, released in November 2023, took the top place among pre-trained base models on Hugging Face's open leaderboard, ahead of Llama 2 70B and Falcon 180B, and the technical report was unusually direct about the reason: not the architecture, which is the ordinary transformer, but the data. Three point one trillion tokens of English and Chinese passed through a cascaded deduplication and quality-filtering pipeline, and a supervised fine-tuning set of fewer than ten thousand dialogues, each revised by hand — a case for curation over volume at a moment when scaling laws were being read as permission to hoover up everything. Sizing the model so that a quantised version ran on a single consumer card widened who could use it at all. Yi also forced the argument about what Chinese laboratories were actually doing. The release reused Llama's architecture with tensors renamed, which abroad was read as copying; the ensuing dispute, in which EleutherAI pointed out that nearly all such models share an architecture and differ in their training data, set the terms for how later Chinese releases were judged. Yi-1.5 followed in 2024 under Apache 2.0, before the company abandoned pre-training and turned to enterprise software.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).