Bespoke Labs, which builds tools to automate the training phase where AI models refine their reasoning and output quality, has secured $40 million（約64億円） in funding. The startup's platform generates reinforcement learning environments using automation and human expertise, then optimizes model performance—a workflow developers typically handle manually and which can be highly time-consuming. The capital will fund platform improvements and further data research.
Summaries like this, in your inbox every morning.
Sign up free →What happened
Bespoke Labs, a startup that streamlines the post-training phase of AI development, raised $40 million（約64億円） in funding across two tranches. A Series A round led by Wing VC brought in $31.75 million（約51億円）, joined by Mayfield, The House Fund, and employees from firms including Anthropic. An earlier tranche of $8.25 million（約13億円） came from a group that included Google DeepMind chief scientist Jeff Dean.
Why it matters
Post-training—the second phase of building custom AI models—typically requires reinforcement learning, a method that involves creating virtual environments and manually iterating on model output. Bespoke's platform automates this workflow using AI-generated simulations and human expert input, claiming to do so significantly faster than traditional manual approaches. This may help AI developers accelerate the often time-consuming process of refining models before deployment.
What to watch
Bespoke will use the new capital to enhance its reinforcement learning platform and fund additional AI data research. The company has also released open-source tools: GEPA for automating prompt engineering and OpenThoughts, a dataset containing more than a million sample prompts and responses for supervised fine-tuning (an alternative post-training method).
Post-training has emerged as a critical bottleneck in AI development because it directly shapes how well a deployed model performs on real-world tasks. The body describes two main post-training methods: reinforcement learning, which trains models on sample tasks in simulated environments, and supervised fine-tuning, which feeds models example prompts and answers. Both are labor-intensive; reinforcement learning requires building custom sandboxes for each use case (a productivity agent needs different simulations than a coding agent), and supervised fine-tuning requires assembling large datasets of question-answer pairs—a process the body calls "highly time-consuming."
Bespoke's approach targets both bottlenecks. Its platform automates the creation of training environments using automation workflows and human expertise, reducing the manual engineering overhead. It then applies automated optimization (via tools like GEPA) to improve model output without requiring engineers to manually experiment with every prompt variation. The funding, which includes backing from a major AI lab (Google DeepMind) and established venture firms, signals confidence that tooling for post-training—once considered routine infrastructure—can now be a standalone venture with substantial investor interest. Bespoke's open-source releases (GEPA and OpenThoughts) appear designed both to build adoption and to validate that its automation approach yields faster, better results than existing manual and dataset-based alternatives.
No discussion yet for this article
Get curated AI news from 200+ sources delivered daily to your inbox. Free to use.
Get Started FreeFree · takes 30 seconds · unsubscribe anytime
1 minute a day. The AI essentials.
200+ sources · Email / LINE / Slack
