Chai AI

Chai AI, also known as Chai Research, builds a consumer platform where users create and share AI characters and stories. Founded by William Beauchamp and based in Palo Alto, it combines user-created personas with conversational models optimized through user feedback. Its developer platform, Chaiverse, hosts submitted models and evaluates them with app users. Chai AI is distinct from Chai Discovery, the biotechnology company.[4][1][2][8]

The character platform grew out of an earlier developer API. Beauchamp’s account of that transition explains how consumer-created personas became central to Chai’s product.[1]

Beauchamp came to conversational AI after running an algorithmic trading company, where he had used neural networks and random forests to predict asset prices. In his January 2025 interview, he described wanting to build a platform on which small teams could develop AI for specialized purposes. Chai’s later company history dates the start of his work with his sister in Cambridge, UK, to 2020.[1][8]

The first product hosted and executed Python scripts with a text-in, text-out interface. Beauchamp recalled building news and recipe bots that attracted little interest. His sister then built a therapist bot; the next day, he observed 20 active users spending an average of 20 minutes in the app, chiefly talking to that bot. He interpreted this as demand for conversation and human-like interaction.[1]

Chai subsequently let consumers create bots using GPT-J, a six-billion-parameter open-source language model, by choosing a prompt, image and name. Beauchamp recalled users creating confrontational characters and romantic archetypes he would not have thought to build himself. These tools broadened the range of experiences beyond his own guesses about what users wanted.[1]

Chai’s July 2025 announcement dates its platform launch to 2021 and reaching one million users to early 2022; its retrospective roadmap separately labels an App Store launch in November 2022. Beauchamp recalled that competition with Character AI exposed the limitations of Chai’s six-billion-parameter model and its smaller compute budget. He said the company moved to Silicon Valley, obtained funding and continued developing its models and feedback systems.[3][4][1]

Chaiverse extends participation from character creation to model development. In Beauchamp’s description, Chai downloads a submitted model, hosts it on a GPU and routes traffic to it. Users compare alternative completions, producing an Elo ranking of which models they find more entertaining. Developers receive scores and other feedback to guide subsequent submissions, with user privacy constraining the information shared.[1]

In January 2025, Beauchamp reported evaluating 20–50 LLMs daily and iterating through at least 100 weekly with roughly five AI researchers. He said Chai had shortened a feedback loop based on waiting 30 days for retention results to roughly three hours. This faster signal supported experiments in fine-tuning, direct preference optimization, prompt engineering, blending and reward models.[1]

Conversation-level blending selects a component model randomly for each response. The selected model receives the conversation history, including replies written by other components, so their different characteristics can influence later turns. Only the selected component generates the next response. This differs from generating outputs from every component and combining them.[5]

The January 23, 2024 version of “Blending Is All You Need” credits Xiaoding Lu, Zongyi Liu, Adian Liusie, Vyas Raina, Vineet Mudupalli, Yuwen Zhang and William Beauchamp, with affiliations spanning Chai Research, the University of Cambridge and University College London. The authors tested a blend of Pygmalion 6B, Chai Model 6B and Vicuna 13B, adapted with conversational fine-tuning and reward-model rejection sampling. They reported higher engagement and retention on Chai than their OpenAI Davinci/GPT3.5 comparator during a 30-day platform experiment. These were measures of conversational use on Chai, rather than general intelligence.[5]

Rejection sampling makes a separate choice among candidate responses. Chai’s roadmap dates Best-of-8 deployment to December 2023. By his January 2025 interview, Beauchamp described generating 16 complete candidates and using a reward model to select a response before displaying it. He linked this approach to Chai’s decision to deliver whole replies without streaming. He also credited the external inference team MK1 for an engine suited to Chai’s workload.[4][1]

Chai’s retrospective roadmap dates its first in-house 6B model to February 2023, model blending to November 2023 and direct preference optimization to March 2024. It records a production blend of 24B models in October 2024 and deployment of group relative policy optimization in March 2025. Later milestones include AMD MI325X and MI300X nodes in September 2025 and a production blend incorporating two 235B A22B mixture-of-experts models in December 2025.[4]

Chai removed its voice feature after testing showed little benefit, according to Beauchamp. He recalled three months spent addressing latency, cost and interface design, followed by an A/B test with no meaningful change in retention, engagement or monetization. Only about 10–15% of users activated audio, and those users used it for about 10–15% of their time. The experiment informed Chai’s decision to focus on other parts of the experience.[1]

Beauchamp also described a 2024 infrastructure migration after Chai’s use of Firebase encountered scaling problems around 500,000 daily active users. The team spent at least three months deciding which requests to retain on Firebase and which to move elsewhere. He attributed outages and interrupted growth to that period. Chai later deployed Model Mesh, its own inference orchestration platform; in July 2025, the company reported serving over 10,000 requests per second across GPU providers and model architectures, alongside more than 1.2 trillion tokens daily.[1][4][3]

On October 2, 2023, Chai announced that it had closed a strategic investment from CoreWeave following a $2 million pre-seed round. The deal carried a $450 million valuation cap; the investment amount was unspecified. Chai said the funding would support its language-model competition and that every platform message then passed through CoreWeave’s cloud.[6]

On July 2, 2025, Chai reported cumulative investment exceeding $55 million and named Ken Howery, Bill Elmore, CoreWeave and AMD among its investors. It reported two million daily active users, over ten million users overall, $40 million in annualized revenue and a team of 11 engineers. Beauchamp said the investment would be spent on doubling compute per user. Chai’s 2026 memo separately states that AMD invested $30 million in 2025.[3][7]

By July 12, 2026, Chai reported a revenue run rate above $100 million a year and a team of around 15 members. The release’s title described the figure as ARR; it did not establish completed-year revenue.[8]

In the same July 2026 release, Chai said its app was undergoing Apple App Store review. Beauchamp described user-generated-content moderation and model safety as separate challenges, with each safety iteration starting with red teaming against the production system. He said the team had introduced a new AI governance system and tools for greater human oversight. His characterization of the version as Chai’s safest was a company assessment; the announcement did not report an App Store approval outcome.[8]