Yi 6B

Yi-6B was 01.AI's debut open model in late 2023, the first serious bilingual Chinese-English release from Kai-Fu Lee's lab, and it made its name by beating Llama 2 7B on most benchmarks despite being smaller. It is a dense Llama-style transformer trained on 3 trillion tokens with particular strength in Chinese: 72.0 on C-Eval and 75.5 on CMMLU in the company's technical report, alongside 63.2 MMLU in English. A 200K-context variant extends reach with NTK-aware RoPE scaling.

Time has moved on and newer 6-7B models now beat it, but Yi-6B remains a solid Apache 2.0 option for Chinese-heavy workloads on modest hardware, with broad GGUF and AWQ quantization support from the community. The base context is only 4K tokens, math and code scores were never its strength, and 01.AI's later Yi 1.5 releases supersede it for most new projects. Still, as a permissively licensed bilingual workhorse, it holds up.

• 01.AI technical report: MMLU 63.2, CMMLU 75.5, C-Eval 72.0, GAOKAO 72.2, BBH 42.8
• Open LLM Leaderboard: HellaSwag 76.4, ARC-Challenge 55.9, Winogrande 73.3, TruthfulQA 42.4 (third-party)