Eugene Cheah

Eugene Cheah is the co-founder and chief executive of Featherless.ai, a serverless inference company that makes thousands of open-weight AI models available without requiring customers to manage their own infrastructure. He also co-leads the RWKV open-source project, developing more efficient alternatives to conventional transformer architectures.

From software reliability to open-model infrastructure

In 2016, Cheah co-founded UI-licious with Shi Ling Tai and became chief technology officer of the Singapore-based automated website-testing company. His writing on software testing describes how automated tests helped stabilize a fragile billing system after rushed releases accumulated technical debt and customer complaints.

He subsequently became a co-lead of RWKV and co-authored research on RWKV-5 and RWKV-6, introducing matrix-valued states and dynamic recurrence for efficient sequence modeling. His work on GoldFinch investigated hybrid RWKV-transformer architectures with compressed key-value caching; a later paper on Key-Value Means, co-authored with Daniel Goldstein, explored reducing cache memory for long-context processing.

Cheah founded Featherless.ai with Harrison Vanderbyl and Wesley George to make open models easier to deploy. The company raised $5 million in seed funding in March 2025 and announced a $20 million Series A in April 2026, co-led by AMD Ventures and Airbus Ventures. By its Series A, Featherless offered more than 30,000 open models and was expanding into enterprise deployments, hardware optimization, agent infrastructure, and regionally controlled computing.

• Production model stability: Enterprises need predictable behavior, manageable costs, permissive licensing, and control over model upgrades. Cheah highlights Mistral NeMo 12B and its Apache 2.0 license as examples of why established models can remain commercially valuable after newer alternatives arrive.

• Human-in-the-loop automation: Insurance and logistics systems can draft responses, consult existing business software, and leave consequential decisions for human review. Automation should expand only after particular workflows demonstrate dependable real-world performance.

• Specialized models over universal benchmarks: Creative writing, companionship, coding, retrieval, and operational workflows demand different capabilities.

• Efficient model architectures: His RWKV, GoldFinch, and Key-Value Means research addresses the memory and inference costs that determine whether open models remain practical to operate at scale.