Hexo Labs
Hexo Labs is a research lab building open-source self-improving AI intended to accelerate superintelligence. Its core thesis is that agents should evaluate their own performance, adapt their strategies, and improve their harnesses and underlying model weights through structured feedback loops.
At a Glance
- AI research labs and researchers
- Enterprise AI teams
- Organizations operating large fleets of agents
- Science, engineering, business, and strategy teams
- +1 more
AI Tools by Hexo Labs
(1)
SIA (Self-Improving AI)
Open Source Self Improving Agent Framework
Discussions
No discussions yet
Be the first to start a discussion about Hexo Labs
Latest News
USA Today feature: Hexo Labs Builds Self-Improving AI That Learns on the Fly
Hexo Labs announced the launch of SIA, an open-source self-improving AI agent, and introduced its Frontier Research Grants program.
SIA: Self Improving AI with Harness & Weight Updates was submitted to arXiv, reporting combined harness-and-weight improvements across legal classification, GPU-kernel optimization, and single-cell RNA denoising.
Hexo published Accelerating superintelligence starts with systems that build themselves, announcing SIA as an open-source recursively improving AI framework.
Products & Services
An open-source framework for recursively improving AI systems. A feedback agent generates hypotheses, evaluates outcomes, and updates both the task agent's harness/scaffold and underlying model weights across generations. The repository is MIT licensed.
A benchmark for measuring whether frontier language models can sustain coherent reasoning across long-horizon chains of thought; it contains approximately 2,500 expert-designed problems across logic, computer science, chemistry, chess, and mathematics.
Hexo's earlier enterprise-focused product: a fine-tuner and open-source image-generation API with controls intended to make generation more predictable and accurate and to help organizations build custom image-generation engines around their design language, products, characters, and IP.
A program offering researchers access to capital, SIA infrastructure, and commercialization opportunities, with selected researchers receiving access to infrastructure and direct collaboration with the Hexo Labs team.
Market Position
Hexo positions SIA as foundational infrastructure for agents that improve rather than remain static, and differentiates it from systems that update only the harness or only model weights by combining both levers. In the earlier image-generation market, contemporaneous reporting identified Stability AI and Regie.ai as global counterparts; in self-improving-agent research, the SIA paper compares against harness-update approaches such as Hyperagents and Meta-Harness and weight-update approaches such as Discover-TTT.
Leadership
Founders
Kunal Bhatia
CEO and co-founder. A three-time founder who previously built SuperLearn, an education company, and Switch, an IoT company; he also worked on AI applications and policy in government and has more than a decade of experience building AI products.
Vignesh Baskaran
CTO and co-founder. A deep-learning and machine-learning practitioner, Top 1% Kaggle performer, and former leader of multiple machine-learning research teams; he has more than a decade of experience building AI products.
Executive Team
Kunal Bhatia
CEO and Co-founder
Serial entrepreneur who previously founded SuperLearn and Switch, and worked on AI applications and policy in government.
Vignesh Baskaran
CTO and Co-founder
Deep-learning practitioner, Top 1% Kaggle performer, and former leader of multiple machine-learning research teams.
Founding Story
Bhatia and Baskaran met during Antler's first Antler India Residency and initially set out to build infrastructure and tooling for enterprise generative AI. After operating a consulting practice that built custom models and agents, they concluded that growth was constrained by the founders' capacity to design and iterate systems; this led to the idea of an AI that could build and improve AI, and ultimately to Hexo's focus on recursive self-improvement.
Business Model
Revenue Model
The earlier Hexo business targeted enterprise revenue from custom image-generation infrastructure/API and fine-tuning services, with business and individual subscription language appearing in Hexo's terms. The current SIA effort is open source and is complemented by research grants and collaboration/commercialization opportunities; public current pricing is not shown.
Target Markets
- AI research labs and researchers
- Enterprise AI teams
- Organizations operating large fleets of agents
- Science, engineering, business, and strategy teams
- Game, film, animation, and VFX companies (for the earlier image-generation product)
- Continuous improvement of production AI agents and fleets of agents
- Scientific research in quantum research, new energy, particle physics, biotechnology, and AI alignment
- Engineering and software optimization
- Business and strategy tasks with measurable outcomes
- Enterprise teams seeking control and ownership of their model and agent stack
- Long-horizon reasoning and benchmark-driven agent research