Yi
Kai-Fu Lee's models, bilingual and open-weight.
Olya's profile
- What sets it apart
- Yi is an open-weight family under the Apache 2.0 license (6B, 9B, 34B), bilingual Chinese-English from pre-training onward. What sets it apart is the mix of genuinely reusable weights, commercial use included, and long-context variants up to 200K tokens. There's also Yi-Lightning, a newer but closed MoE model available only via API.
- Strengths
- The open weights run locally with no licensing strings attached, and the 34B holds up well on reasoning and instructions for its size; Yi-Coder covers 52 programming languages with a 128K context. On text that mixes Chinese and English it's among the best for its size, and Yi-Lightning's API pricing (about $0.14 per million tokens) is aggressive.
- When to use it
- It fits when you work with bilingual Chinese-English content and want weights you can host yourself, commercial use included, without depending on a vendor. The 9B or 34B is a reasonable base to fine-tune, and Yi-Lightning via API suits cost- and latency-sensitive applications.
- When to avoid it
- Since March 2025 01.AI has stopped pre-training its own models from scratch, shifting to enterprise services: the open Yi weights are effectively frozen at 2024, so for frontier capability you're better off looking elsewhere. If you need broad multilingual support beyond Chinese-English, or commercial guarantees on the model, the alternatives are more solid.