# Bespoke Labs

> Bespoke Labs is a Mountain View AI research lab that builds tools and datasets for training more reliable AI models and agents, starting with synthetic data curation and hallucination detection and expanding into realistic reinforcement-learning environments for long-horizon agentic AI. Its open-source projects, including Bespoke Curator, Bespoke-MiniCheck, OpenThoughts, Terminal-Bench, and GEPA, are widely used across the AI research community, and the company raised $40M in seed and Series A funding by mid-2026 from investors including Wing Venture Capital, 8VC, and Mayfield.

- **Founded:** 2024
- **Headquarters:** Mountain View, California, United States
- **Founders:** Mahesh Sathiamoorthy (Co-founder and CEO), Alex (Georgios) Dimakis (Co-founder and Chief Scientist)
- **Team size:** 42
- **Products:** Bespoke Curator, Bespoke-MiniCheck-7B, OpenThoughts, Terminal-Bench, GEPA (Genetic-Pareto)
- **Notable:** Bespoke-MiniCheck-7B topped the public LLM-AggreFact Grounded Factuality Leaderboard, beating GPT-4o and Claude 3.5 Sonnet, OpenThoughts-114k became the #1 trending dataset on Hugging Face (February 2025), OpenThoughts3-1.2M became the #1 trending dataset on Hugging Face again (June 2025)

## Products & services

- **Bespoke Curator** — Open-source Python library for synthetic data generation, curation, and structured data extraction, with support for retrieval-augmented fine-tuning (RAFT), async batch inference, caching, and data quality indicators.
- **Bespoke-MiniCheck-7B** — A compact grounded-factuality/hallucination-detection model fine-tuned from InternLM2.5-7B-chat; tops the public LLM-AggreFact (Grounded Factuality) Leaderboard, outperforming larger models like GPT-4o and Claude 3.5 Sonnet at roughly 200ms latency.
- **OpenThoughts** — A collaboration with the DataComp community to produce open reasoning datasets (OpenThoughts-114k, OpenThoughts3-1.2M) and models (OpenThinker-7B), downloaded over 500,000 times and used by Meta, Amazon, and AI2.
- **Terminal-Bench** — A benchmark for evaluating agentic coding and terminal-use capability, used by Anthropic, OpenAI, and Google DeepMind to assess frontier model performance.
- **GEPA (Genetic-Pareto)** — An open-source evolutionary, reflection-based prompt and policy optimizer that improves on reinforcement-learning baselines like GRPO using far fewer rollouts; integrated with DSPy.
- **RL Environments Platform** — Commercial infrastructure for composing realistic, simulated agent environments (codebases, microservices, sandboxed execution) to train and evaluate reliable long-horizon AI agents.

## Achievements

- Bespoke-MiniCheck-7B topped the public LLM-AggreFact Grounded Factuality Leaderboard, beating GPT-4o and Claude 3.5 Sonnet
- OpenThoughts-114k became the #1 trending dataset on Hugging Face (February 2025)
- OpenThoughts3-1.2M became the #1 trending dataset on Hugging Face again (June 2025)
- Terminal-Bench adopted by Anthropic, OpenAI, and Google DeepMind for agentic model evaluation
- Raised $40M in combined seed and Series A funding, announced July 2026

## Latest updates

- **2026-07** — Announced $40M in combined seed and Series A funding, led by Wing Venture Capital, to build RL environments for training reliable AI agents.
- **2025-06** — Released OpenThoughts3-1.2M, which became the #1 trending dataset on Hugging Face.
- **2025-01** — Released OpenThoughts-114k dataset and OpenThinker-7B model with the DataComp community.

## Links

- Website: https://bespokelabs.ai
- LinkedIn: http://www.linkedin.com/company/bespokelabsai
- Twitter/X: https://x.com/bespokelabsai
- GitHub: https://github.com/bespokelabsai

---

Profile page: https://yespress.io/bespoke-labs
Published by YesPress — https://yespress.io
Last updated: 2026-07-13