Chai Research

Chai Research builds a consumer platform for creating and interacting with AI characters and stories. Founded by William Beauchamp, it combines user-created content with in-house language models and Chaiverse, a developer platform that evaluates submitted models through consumer feedback. Its research and engineering focus on making conversations engaging while managing the cost of serving many responses.[3][1]

The character platform grew out of an earlier developer API. Beauchamp's initial experiments explain the transition from hosting programmers' bots to giving consumers tools to create their own conversational experiences.[1]

Before Chai, Beauchamp ran an algorithmic trading business, using machine learning to predict asset prices. In his January 2025 interview, he recalled deciding to build a platform where small teams could develop language models for different purposes. He continued running the trading company alongside Chai before exiting it in the third quarter of 2024.[1]

Chai's first product hosted and executed Python scripts with a text-in, text-out interface. Beauchamp built news and recipe bots but recalled little demand for them. A therapist-bot experiment involving his sister produced a different response: he reported finding 20 active users spending an average of 20 minutes in the app. He interpreted that engagement as demand for open-ended conversation.[1]

The next consequential change was letting consumers create bots themselves. Using GPT-J, a six-billion-parameter open-source model, Chai allowed users to choose a prompt, image, and name. Users created argumentative characters and romantic archetypes that Beauchamp said he would not have anticipated. This broadened the content beyond the bots he could invent personally.[1]

Chai dates its launch to 2021 in its financing announcement; its website separately lists an App Store launch in November 2022. Beauchamp estimated that he personally invested approximately £2 million before the company moved to Silicon Valley and obtained additional funding. He described the move as a response to competition from Character AI and the resources needed to serve better models.[4][3][1]

The March 2023 preprint Rewarding Chatbots for Real-World Engagement with Millions of Users describes training a reward model from automatic pseudo-labels derived from user interactions. At inference time, that model rejects low-scoring candidate responses. The authors use mean conversation length as an engagement proxy and report A/B tests on groups of 10,000 new daily chatbot users: for GPT-J 6B, mean conversation length increased by up to 70%, with user retention increasing by more than 30%. These results measure engagement and retention in the experiments.[6]

The paper's fourteen authors are Robert Irvine, Douglas Boubert, Vyas Raina, Adian Liusie, Ziyi Zhu, Vineet Mudupalli, Aliaksei Korshuk, Zongyi Liu, Fritz Cremer, Valentin Assassi, Christie-Carol Beauchamp, Xiaoding Lu, Thomas Rialan, and William Beauchamp. It proposes future work alternating updates to the language model and reward model using new conversations.[6]

Blending Is All You Need, first submitted in January 2024, investigates combining smaller chat models through thirty-day A/B testing on Chai. Its authors are Xiaoding Lu, Zongyi Liu, Adian Liusie, Vyas Raina, Vineet Mudupalli, Yuwen Zhang, and William Beauchamp. In the later interview, Beauchamp illustrated blending with a simple example: randomly serve half the requests to a smart model and half to a funny model, giving a conversation access to both models' strengths.[5][1]

Chai's deployment timeline places its first in-house 6B model in February 2023, Best-of-4 reward model in March, Proximal Policy Optimization in May, model blending in November, and Best-of-8 rejection sampling in December. It dates Direct Preference Optimization to March 2024 and a production blend of ten 24B models to October 2024.[3]

Chaiverse hosts developers' submitted models on GPUs, routes consumer traffic to them, and ranks responses through users' comparisons. In January 2025, Beauchamp estimated that approximately 5,000 completions supplied a useful ranking signal for entertainment. He reported evaluating 20–50 models daily and iterating through at least 100 weekly, with a feedback loop shortened from thirty days to approximately three hours. Developers received scores to guide further submissions.[1]

Beauchamp said Chai had paid model creators more than $100,000. He recalled a roughly seventeen-year-old recipient spending a $1,000 payment on a physical computer for training more models. Chai also offered no-strings-attached grants for open-source projects, which he described as both community support and a way to meet talented developers.[1]

Growth exposed limits in Chai's backend. Beauchamp recalled difficulties with its use of Firebase around 500,000 daily active users in 2024, followed by at least three months migrating requests among services. Outages interrupted the user experience and, in his account, hurt retention and organic app-store acquisition.[1]

Chai also removed a voice feature after investing three months in its development. Beauchamp said its A/B test showed no change in key metrics; only about 10–15% of users activated audio, and they used it for about 10–15% of their time. The result reinforced his emphasis on features affecting a substantial share of the conversational experience.[1]

Beauchamp credited MK1 with an inference-engine collaboration after Chai had used vLLM. He described generating sixteen candidate completions and evaluating the complete responses with a reward model before displaying the selected answer. This selection process explains Chai's decision to show responses all at once rather than stream tokens. He described predicting whether a user would reply as one possible reward-model target.[1]

The January 26, 2025 interview reported approximately 1.4 million daily active users and more than $22 million in revenue, without specifying the revenue period. Beauchamp said users had tripled during 2024 and revenue had more than doubled.[1]

On July 2, 2025, Chai announced that it had secured additional financing that brought cumulative investment above $55 million. It named Ken Howery, Bill Elmore, CoreWeave, and AMD as investors. Beauchamp stated the spending plan: “This investment will be spent on doubling the amount of compute per user, empowering us to serve cutting-edge AI to our users.”[4]

The same announcement reported more than 10 million users, separately identifying 2 million daily active users, an eleven-engineer team, and $40 million ARR, also described as annualized revenue. Chai reported serving more than 1.2 trillion tokens daily. Its custom Model Mesh orchestration platform supported different GPU providers and model architectures, with reported throughput above 10,000 requests per second.[4]

Chai dates Model Mesh's deployment to January 2025 and Group Relative Policy Optimization to March 2025. Its later timeline records AMD MI325X and MI300X deployment in September, mixture-of-experts training beginning in November, and two 235B A22B models entering the production blend in December. These deployments extended its model-serving infrastructure across GPU types and larger model architectures.[3]