Bespoke Labs Raises $40 Million For AI Post-Training Tools

Bespoke Labs said it raised $40 million to expand software for AI post-training, the stage in which developers refine a model after pre-training and before production use.

The company said the money will support its reinforcement learning platform and additional AI data research.

The financing came in two tranches.

Bespoke said it raised $31.75 million in a Series A round led by Wing VC, with Mayfield, The House Fund and employees at major technology firms such as Anthropic also joining.

Bespoke said it previously raised $8.25 million from a group that included Google DeepMind chief scientist Jeff Dean.

The disclosed round puts the startup inside a technical part of the AI stack that is becoming more important as companies try to turn base models into agents, coding tools and enterprise workflows.

Bespoke's own claims remain company-owned: the announcement did not include customer names, revenue, valuation or third-party benchmark results.

Post-Training Platform Targets Reinforcement Learning Environments

Bespoke's platform is built around the post-training work that follows pre-training.

Pre-training gives a neural network broad capabilities, while post-training is used to sharpen reasoning and task performance for narrower use cases.

Developers often use reinforcement learning for that step.

In that process, an AI model receives sample tasks similar to the work it is expected to do, then receives a reward signal when it completes the task correctly.

The reward data changes the model's configuration and is intended to improve output quality.

The company said its software helps create the virtual environments used for reinforcement learning.

A productivity agent might need a sandbox that looks like an employee workstation, while a coding agent might need a simulated GitHub repository.

Bespoke said its platform generates those simulations through automation workflows and input from a network of human experts.

Bespoke also said its platform uses a sandboxing layer to run the generated AI environments.

The company claims that layer helps minimise latency and increase throughput, but it did not publish independent benchmark methodology or customer validation for those performance claims.

GEPA And OpenThoughts Show The Open-Source Data Push

Bespoke is also using open-source projects to support its post-training position.

The platform includes GEPA, an open-source project the company released last year to automate prompt engineering.

That work focuses on finding the requests and prompt formats that maximise model output quality.

The company is working on supervised fine-tuning as well as reinforcement learning.

Supervised fine-tuning gives AI models sample prompts and answers that they can use to refine outputs.

Bespoke released a dataset called OpenThoughts last January.

The company said the dataset contains more than a million sample prompts and responses, and said it produces better post-training results than earlier SFT datasets.

Those dataset claims come from Bespoke; the funding announcement did not include an independent comparison table or named external deployment using OpenThoughts.

Funding Leaves Customer And Valuation Evidence Undisclosed

The new capital gives Bespoke more room to build in a crowded AI tooling market, but the available evidence is mainly funding size, investor list and product description.

The company named the round size and several backers, and it described how its reinforcement learning environments, sandboxing layer, GEPA work and OpenThoughts dataset fit into its platform.

The announcement did not name paying customers, disclose annual revenue, state a valuation, publish deployment counts, provide third-party latency or throughput tests, or identify enterprise teams using the platform in production.