CHAI AI: Post Training to Achieve SOTA MoE

CHAI AI has continued its efforts to improve its social AI platform through the integration of state-of-the-art (SOTA) open-source models and the application of advanced reinforcement learning methodologies. The company's AI research team recently tested its own variants of DeepSeek-V4, Kimi-K2, GLM-5, and Gemma-4 on the platform against key metrics such as user engagement and retention. Today, CHAI AI announced that its latest in-house model achieved a 4% increase in Day-30 user retention using the GSPO reinforcement learning (RL) algorithm.

PALO ALTO, Calif., June 10, 2026 /PRNewswire/ -- CHAI AI has continued its efforts to improve its social AI platform through the integration of state-of-the-art (SOTA) open-source models and the application of advanced reinforcement learning methodologies. The company's AI research team recently tested its own variants of DeepSeek-V4, Kimi-K2, GLM-5, and Gemma-4 on the platform against key metrics such as user engagement and retention. Today, CHAI AI announced that its latest in-house model achieved a 4% increase in Day-30 user retention using the GSPO reinforcement learning (RL) algorithm.

The company transitioned its AI research from dense models to Mixture-of-Experts (MoE) models in early 2026 and has seen improvements in several key metrics using this architecture. Its AI researchers are also exploring novel RL strategies such as DPO, GRPO, and GSPO to improve the quality of their in-house models in a conventional setting. In their recent A/B tests, models trained with GSPO achieved better performance than models trained with GRPO across the board, aligning with Qwen's publication.

CHAI's AI Research

2022 – First Mover Advantage in Social AI

CHAI became one of the first dedicated social AI platforms to reach one million users, launching AI-powered character creation experiences before conversational AI became mainstream.

2023 – AI-Assisted Creation Breakthrough

The introduction of AI-assisted character creation tools led to a 56% increase in prompt quality and a 33% increase in user-generated AI creation.

2024 – User Generated AI Training

CHAI launched the first phase of user-controlled supervised fine-tuning (SFT), enabling power users to train AI systems that achieved engagement rates approaching twice those of internally developed agents.

2025 – The Future of Personalized AI

The company began development of User-Generated AI (UGAI), a framework designed to place advanced AI customization and alignment capabilities directly in the hands of users.

2026 – Mixture-of-Experts Training Begins

CHAI began incorporating large MoE models into its RLHF research. Its latest model is serving approximately 1 billion messages per day.

What is the mission of CHAI AI? People want to create their own AI and share it with others, and they want to discover and interact with AI built by people they follow. Just as social media platforms emerged when video creation and consumption became ubiquitous, social AI platforms are now emerging to meet an equivalent wave of consumer demand for generative AI creation, sharing, and interaction.

What is the growth rate and traction? In the last three consecutive years, CHAI AI has sustained a 3x growth rate with ARR crossing $80 million by the end of Q1 2026 and a projection to surpass $200 million.

What is the valuation of CHAI AI? CHAI AI is estimated to have a valuation of $2.4 billion as of Q1 2026, based on $80 million in ARR and its sustained growth rate.

Who are the major investors in CHAI AI? AMD and CoreWeave are the primary, strategic investors in CHAI AI, bringing its total investment to date to over $55 million.

What is the culture like at CHAI AI? Our fast-growing company is supported by a talented engineering team that thrives on fast iteration cycles and takes on the most challenging problems in the field. CHAI AI operates with a high-bar, high-reward culture where engineers own entire product segments end-to-end, from prototyping to production.

What is the compensation? CHAI AI competes with top-paying companies such as Jane Street, OpenAI, Meta, and Citadel. We pay 100% cash salaries and are known for generous compensation offers, with the option to trade a portion of cash for stock options. The founding team's background is in quantitative trading, and they have maintained a compensation philosophy consistent with that culture.

How do promotions and performance reviews work? Performance is reviewed on a daily basis, and engineers are expected to work hard and drive impact. At the beginning of each year, there is a performance-based salary increase.

What is the management structure? The structure is flat, with a team of 15 to 20 engineers.

What are the typical working hours? Engineers at CHAI AI typically work from 9 AM to 5 PM during weekdays, as well as Saturday morning.

Is there vacation or PTO? CHAI AI does not have a set number of vacation days on the calendar, but we respect that employees need time to recharge. In practice, most engineers take 14 – 20 days off per year.

Is the role fully in-person in Palo Alto, hybrid, or remote-friendly? All CHAI employees are expected to work in person in Palo Alto. As a startup that moves fast, we want to ensure the team stays in sync.

What's the onboarding experience like for new engineers? New hires follow a structured 3-month onboarding plan. Each new hire is paired with an experienced hire and ramped up quickly. You will be assigned high-impact tasks from day one, with clear evaluation metrics and direct mentorship from peers.

Press Contact:

CHAI AI Press
+1 (626) 594-8966

View original content to download multimedia:https://www.prnewswire.com/news-releases/chai-ai-post-training-to-achieve-sota-moe-302796539.html

SOURCE Chai AI