Thinking Machines Lab is an artificial intelligence research and product company focused on building AI systems that are more widely understood, customizable, and generally capable. Its public mission combines frontier model development with a strong emphasis on human-AI collaboration, multimodal interaction, and research that can support real-world scientific and engineering work.
Thinking Machines Lab positions itself as both a lab and a builder. It pairs research with product development, publishes technical work and code through Connectionism, and develops infrastructure intended to make advanced model training and customization more accessible beyond a small set of elite labs. That combination gives Thinking Machines Lab a profile that blends research rigor, developer tooling, and applied AI platform ambitions.
Offerings, Capabilities, and Integrations
Thinking Machines Lab offers AI services built around model training, fine-tuning, inference, evaluation, and hosting. Its capabilities span supervised fine-tuning, reinforcement learning, preference optimization, distillation, sampling, checkpointing, and model weight management, with support for both text and vision workflows. The company emphasizes user control over data, algorithms, and model choice while it handles distributed infrastructure and operational complexity.
Thinking Machines Lab also supports a research-oriented workflow through documentation, interactive tutorials, open-source examples, and programs that encourage broader community participation. Its service model is designed to accommodate third-party models, customer models, and company-controlled models, and it provides interfaces that fit into existing development environments rather than forcing users into a closed stack.
Products and Services
- Tinker: Managed training API for researchers and developers to fine-tune open-weight models with LoRA while Thinking Machines Lab handles distributed compute, scheduling, and infrastructure operations. It supports training, sampling, checkpointing, evaluation, and model weight export across text and vision use cases.
- Tinker Cookbook: Open-source companion library for Tinker that provides realistic post-training examples, higher-level abstractions, deployment workflows, and interactive tutorials for tasks such as supervised fine-tuning, reinforcement learning, distillation, preference optimization, tool use, and multi-agent training.
- Connectionism: Thinking Machines Lab’s research publishing hub for technical posts and shared science, used to release methods, experiments, and other research outputs from the team and collaborators.
- Tinker Console: Self-service interface associated with Tinker for account access, API key management, and viewing current model and pricing information.
Target Customers
Thinking Machines Lab primarily targets AI researchers, ML engineers, and developers who want direct control over post-training methods and model behavior without building or operating their own distributed training stack. Its tooling is aimed at users working with open-weight models and at teams that need to customize models for specific datasets, tasks, or workflows.
The company also targets academic labs, instructors, students, and nonprofit research groups through grant-supported access and open technical materials. In parallel, its service terms and infrastructure posture indicate a fit for organizations and enterprise teams that want to host, fine-tune, evaluate, or deploy models for internal business and production-oriented use cases.
Cloud Integrations and Marketplace
- Google Cloud: Tinker Cookbook supports Google Cloud Storage as a backend for training and evaluation data via gs:// URIs.
- AWS: Tinker Cookbook supports Amazon S3 as a backend for training and evaluation data via s3:// URIs.
- Microsoft Azure: Tinker Cookbook supports Azure Blob Storage as a backend for training and evaluation data via az:// URIs.
Key People
- Mira Murati: CEO
- John Schulman: Chief Scientist
- Soumith Chintala: CTO
- Jonathan Lachman: Founding Head of Operations
- Alec Radford: Advisor
- Bob McGrew: Advisor
Key Facts
- Headquarters: San Francisco, California, United States
- Employees: Approximately 120
- Annual Revenue: Undisclosed
- Parent Company: None
- Subsidiaries: None
- Publicly Listed: No (privately held)
