Yi

Yi (meaning "one" in Chinese, stylised as Yi or 易) is a family of open-source large language models developed by 01.AI (零一万物, Língyī Wànwù), a Beijing-based AI startup founded in March 2023 by Kai-Fu Lee. Lee is a prominent figure in the global AI industry, having led AI divisions at Apple, SGI, Microsoft, and Google before founding the venture capital firm Sinovation Ventures. 01.AI was one of several well-capitalised Chinese AI startups founded rapidly in 2023 following the public launch of ChatGPT.

Founding and Positioning

01.AI was founded with the explicit goal of competing in the global frontier of large language model development, distinguishing itself from many Chinese AI startups by prioritising international-standard open-weight model releases alongside domestic applications. Kai-Fu Lee's international reputation and English-language communication provided the company visibility in Western AI communities that Chinese-language-only startups typically lacked.

The company reached a unicorn valuation (exceeding USD 1 billion) within months of its founding, backed by investors including Alibaba and leading Chinese and international venture capital firms.

Yi-34B and Early Releases

In November 2023, 01.AI released Yi-34B, a 34-billion-parameter model supporting Chinese and English that outperformed significantly larger models including Llama 2-70B and Falcon-180B on standard benchmarks at the time of release. Yi-34B demonstrated that aggressive pretraining on high-quality bilingual data could produce competitive models at a fraction of the parameter count of competing open-weight releases.

01.AI subsequently open-sourced Yi-9B, positioning it as particularly strong for code and mathematical tasks within the same model family. A Yi-6B variant provided a smaller, more hardware-accessible option. The release of Yi models on Hugging Face attracted widespread downloads and fine-tuning experiments from the global open-source AI community.

Yi-1.5 and Yi-Lightning

Yi-1.5 (May 2024) introduced an updated series with improved instruction following and conversational capabilities, incorporating lessons from RLHF alignment and expanding context window support. The series included 6B, 9B, and 34B parameter variants.

Yi-Lightning, released October 2024, was 01.AI's flagship model at the time of its release. In the LMSYS Chatbot Arena leaderboard — which ranks models based on human pairwise preferences — Yi-Lightning placed joint third among all AI companies globally, alongside xAI's Grok-2, behind only OpenAI and Google. This represented a significant achievement for a Chinese startup in a benchmark designed around human preference rather than automated metrics.

Strategic Pivot

In March 2025, 01.AI announced a strategic pivot away from pretraining its own frontier LLMs. The company cited the commoditisation of foundation model training — driven largely by DeepSeek's release of highly capable open-weight models — as making independent pretraining economically unjustifiable for a startup of its scale. Instead, 01.AI repositioned as a company selling tailored AI business solutions and applications built on top of DeepSeek's open-weight models, optimising for enterprise deployment and industry-specific fine-tuning rather than frontier model research.

The Yi model family attracted adoption among Malaysian AI developers and researchers primarily through Hugging Face, where its open-weight releases were accessible for download and fine-tuning without licensing restrictions. Malaysian universities with AI research groups — including Universiti Malaya, UTM, and UPM — downloaded and evaluated Yi models for tasks including Bahasa Malaysia NLP, code generation, and question answering.

Yi's strong Chinese-English bilingual performance made it particularly relevant for Malaysian use cases requiring seamless switching between Chinese and English — such as customer service bots for Malaysian Chinese businesses, content generation tools for Chinese-language media, and educational tools for Chinese independent schools.

The Yi-6B and Yi-9B variants, deployable on NVIDIA RTX 3090 or A100 GPUs commonly available in Malaysian university research labs, enabled practical experimentation without cloud compute costs. Several Malaysian AI startups used Yi as a base model for domain-specific fine-tuning on Bahasa Malaysia or code-switching (Singlish-influenced Malaysian English) datasets.

01.AI's strategic pivot to DeepSeek-based solutions in 2025 is also significant for Malaysia: it reflects the broader trend of Chinese AI ecosystem consolidation around highly capable open-weight base models, with specialised players focusing on application-layer value rather than foundation model training. Malaysian enterprise AI buyers evaluating Chinese LLM solutions in 2025-2026 are increasingly likely to encounter DeepSeek-based products (including those from 01.AI's new enterprise business) rather than models trained independently by individual Chinese AI startups.

MDEC's Digital Skills for All programme and HRD Corp-accredited AI training providers have incorporated open-weight Chinese model evaluation — including Yi variants — into their advanced AI practitioner curricula, reflecting the growing relevance of non-Western model ecosystems for Malaysian AI practitioners.

See Also

References

1. Wikipedia. (2025). 01.AI. en.wikipedia.org.
2. TechCrunch. (2023). Valued at $1B, Kai-Fu Lee's LLM startup unveils open source model. techcrunch.com.
3. VentureBeat. (2023). Chinese AI unicorn's 34B LLM outperforms larger Llama 2 and Falcon models. venturebeat.com.
4. LabLab.ai. (2025). 01.AI's Yi series large language models. lablab.ai/tech/yi-llms.
5. Hugging Face. (2025). 01-ai model collection. huggingface.co/01-ai.