Yi-Lightning
by 01.AI·China·Released
01.AI's fast tier — top-tier quality at consumer-app latency.
About this model
Yi-Lightning (October 2024) is 01.AI's fast tier — Kai-Fu Lee's lab's bet that the future of consumer AI is cheap, fast inference rather than ever-larger frontier models. At ~$0.14/M for both input and output it's one of the cheapest production APIs in the world, while reaching a Chatbot Arena Elo of 1287 (above several closed Western models in the same window).
01.AI doesn't disclose architecture or parameter count for Yi-Lightning. The model is closed-weights via 01.AI's lingyiwanwu.com platform; the lab's earlier Yi-34B and Yi-1.5 series remain open. Particularly strong for Chinese-language workloads where Western frontier models have noticeable quality gaps.
Strengths
- •Among the cheapest production APIs globally — $0.14/M flat
- •Strong Chinese-language performance
- •Surprisingly competitive Chatbot Arena placement for the price
- •Backed by Kai-Fu Lee's long track record
Limitations
- •Closed weights (unlike Yi-1.5 / Yi-34B open releases)
- •16K context — smaller than Western frontier
- •Less mature international developer ecosystem
When to use it
- →High-volume Chinese consumer / enterprise chat
- →Cost-optimised classification and routing
- →Hybrid architectures: Yi-Lightning for the bulk, larger models for edge cases
Architecture & training
Architecture details undisclosed. 01.AI has publicly emphasised data quality + careful curation as their primary lever (consistent with Kai-Fu Lee's stated philosophy on efficient training).
Benchmarks
| Benchmark | Score | Bar |
|---|---|---|
| Chatbot Arena | 1287.0 |