Yi Large

À propos de 01.AI (Lingyi Wanwu)

Fondée 2023 · Beijing, China

01.AI (零一万物, Lingyi Wanwu) was founded in March 2023 by Kai-Fu Lee — former president of Google China, founder of Sinovation Ventures and a globally recognised AI investor and author. The lab launched the Yi-6B and Yi-34B open-weights base models in November 2023, which briefly led the Hugging Face Open LLM Leaderboard. 01.AI followed with Yi-9B (2024), Yi-VL (vision, 2024) and the closed-API flagship Yi Large in May 2024. The company raised $200M+ in seed funding in 2023 at a reported $1B valuation, with investors including Alibaba Cloud, Sinovation Ventures and others. 01.AI's positioning emphasises strong bilingual Chinese-English performance, data-quality-focused training and pragmatic product applications.

Architecture

Decoder-only Transformer (proprietary)

Yi Large is a dense decoder-only transformer following the modern LLM recipe — grouped-query attention, RoPE positional embeddings, SwiGLU activations and RMSNorm. 01.AI has been less specific about Yi Large's exact size than about the open Yi-34B/9B/6B variants. The Yi technical report describes the training methodology in detail: a heavily filtered bilingual Chinese-English data pipeline using CCNet-style quality filtering, aggressive deduplication and curated additions of books, code and scientific content. The Yi-34B base was trained on 3.1T tokens; Yi Large is reported to scale further. Post-training is supervised fine-tuning plus RLHF with proprietary specifics. Yi Large is closed-API only — only the smaller Yi base and chat variants (Yi-34B, Yi-9B, Yi-6B) have open weights under the Yi Series Models Community License Agreement. Released May 2024 via 01.AI Platform and partner clouds.

Paramètres

Undisclosed (estimated ~100B based on tier sibling Yi-34B and contemporaneous public details)

Contexte

32 000 tokens

Capacités

• Top-tier bilingual Chinese-English performance
• Frequent #1 or #2 ranking on SuperCLUE Chinese leaderboards
• Top-10 LMSYS Chatbot Arena ranking at release in 2024
• Strong Chinese reasoning, math, writing and exam-style tasks
• Claimed parity with GPT-4 on selected English benchmarks at release
• 32K context window
• Available via 01.AI Platform and hosted API
• Best for: Chinese-market applications, bilingual content workflows, Chinese exam and education AI.

Entraînement et licence

Trained on heavily filtered bilingual Chinese-English web data, books, code and academic content. Smaller Yi-34B base used 3.1T tokens; Yi Large is reported to scale further. Data pipeline uses CCNet-style quality filtering plus aggressive deduplication (described in the Yi technical report). Knowledge cutoff approximately early 2024. Post-training is supervised fine-tuning plus RLHF.

Licence: Yi Large itself is closed-API only under 01.AI's commercial terms of service. The open Yi base / chat variants (Yi-34B, Yi-9B, Yi-6B) use the Yi Series Models Community License Agreement, which permits commercial use after free registration with 01.AI.

Tests de sécurité: 01.AI publishes model cards for the open Yi variants. Yi Large's safety profile aligns with Chinese regulatory requirements — the model filters politically sensitive topics consistent with Chinese AI regulations. No third-party red-team report.