Mira Murati’s Thinking Machines Lab launches Tinker, a $12B AI startup tool simplifying LLM fine-tuning and democratizing AI customization.

In the rapidly evolving landscape of artificial intelligence, innovation often stems from visionary leaders who challenge the status quo. On October 1, 2025, Thinking Machines Lab, founded by former OpenAI Chief Technology Officer Mira Murati, unveiled its inaugural product, Tinker. This Python-based API and managed service is designed to simplify the fine-tuning of large language models (LLMs), making advanced AI customization accessible to researchers, developers, and startups alike. As AI continues to permeate industries from healthcare to finance, tools like Tinker could democratize access to frontier technologies, reducing the barriers traditionally imposed by resource-intensive processes.

This launch marks a significant milestone for Thinking Machines Lab, a startup valued at $12 billion following a record-breaking $2 billion seed funding round led by Andreessen Horowitz, with participation from NVIDIA, Accel, and other prominent investors. Murati, who played a pivotal role in developing groundbreaking products at OpenAI such as ChatGPT, left the company in September 2024 to pursue this venture. Her departure, amid internal shifts at OpenAI, underscored a growing trend of AI pioneers branching out to foster more open and collaborative ecosystems.

The Genesis of Thinking Machines Lab

Mira Murati’s journey in AI is nothing short of remarkable. At just 36 years old, she has already left an indelible mark on the field. Prior to her tenure at OpenAI, where she served as CTO and briefly as interim CEO during a high-profile leadership transition in 2023, Murati held key positions at Tesla and Leap Motion, honing her expertise in engineering and AI applications. Her vision for Thinking Machines Lab, established in February 2025, centers on bridging critical gaps in AI development: enhancing scientific understanding of frontier systems, decentralizing training knowledge from elite labs, and enabling seamless customization of AI models to align with diverse user needs and values.

The company’s ethos emphasizes shared science, inclusive AI, robust foundations, and empirical learning. By collaborating with the broader research community through publications, technical blogs, and open-source contributions, Thinking Machines Lab aims to accelerate collective progress in AI. This approach contrasts with the more proprietary models of giants like OpenAI, promoting transparency and accessibility in an industry often criticized for its opacity. Murati’s team, comprising alumni from OpenAI, Meta, and other leading firms, brings experience from projects like PyTorch, Fairseq, and Segment Anything, ensuring a blend of rigorous engineering and creative innovation.

In an era where AI ethics and safety are paramount, Thinking Machines Lab commits to high standards, including iterative safety measures and support for external alignment research. This proactive stance addresses concerns about AI’s potential misuse, positioning the lab as a responsible player in the ongoing “AI arms race.”

Introducing Tinker: A Tool for AI Customization

At the heart of the launch is Tinker, an API that streamlines the fine-tuning of open-weight language models. Fine-tuning involves adapting pre-trained models to specific tasks or datasets, a process that has historically required substantial computational resources and expertise. Tinker abstracts these complexities, allowing users to write simple Python scripts on their local machines and submit jobs for distributed training on the lab’s GPU clusters.

Supporting models from Meta’s Llama series to Alibaba’s Qwen (up to massive variants like Qwen-235B), Tinker facilitates supervised fine-tuning, reinforcement learning (RL), and custom experiments. Techniques such as Low-Rank Adaptation (LoRA) enable efficient training by allowing multiple tasks to share computational resources, significantly lowering costs and barriers. Users retain full control over their data and algorithms, with the platform handling resource allocation, fault recovery, and scaling automatically.

Early beta testers, including researchers from Princeton, Stanford, Berkeley, and Redwood Research, have praised Tinker’s simplicity. For instance, one user highlighted its ability to embed backdoors in code generation models through RL, a task made far less cumbersome than traditional methods. The accompanying open-source Tinker Cookbook provides templates for common workflows, such as prompt distillation and retrieval-augmented generation, further empowering developers to focus on innovation rather than infrastructure.

Key Features and Technical Edge

Tinker’s appeal lies in its user-centric design. Key features include:

• Seamless API Integration: Developers can prototype locally and scale effortlessly to cloud-based clusters, supporting models of varying scales, including mixture-of-experts (MoE) architectures.
• Efficiency with LoRA: By leveraging parameter-efficient methods, Tinker reduces the computational footprint, making fine-tuning viable for smaller teams without access to vast data centers.
• Customization Flexibility: From math reasoning to multi-agent systems, users can tailor models for specialized applications like theorem proving or chemistry simulations.
• Export and Deployment: Fine-tuned models can be downloaded and deployed in any environment, ensuring portability and ownership.
• Safety and Control: While empowering experimentation, the platform encourages responsible use, aligning with the lab’s commitment to AI safety.

This combination positions Tinker as a bridge between academic research and practical deployment, potentially accelerating advancements in fields requiring hyper-specialized AI.

The Launch: Private Beta and Future Plans

Announced via the company’s blog and Murati’s X post, Tinker’s launch has generated buzz in tech circles. Currently in private beta and offered for free, the service is accessible via a waitlist at thinkingmachines.ai. Usage-based pricing is slated for introduction soon, reflecting the high costs of GPU infrastructure. The rollout emphasizes collaboration, with initial users from top universities testing its capabilities in real-world scenarios.

Today we launched Tinker.

Tinker brings frontier tools to researchers, offering clean abstractions for writing experiments and training pipelines while handling distributed training complexity. It enables novel research, custom models, and solid baselines.

Excited to see what… https://t.co/rLSluOckuC

— Mira Murati (@miramurati) October 1, 2025

Reactions on Social Media platform  highlight excitement about democratizing AI tools, though some note it’s “useful but not blockbuster,” suggesting it builds on existing frameworks rather than introducing radical innovations. Nonetheless, its focus on open-source models challenges closed systems, potentially shifting power dynamics in the AI ecosystem.

 Andrej Karpathy (Previously Director of AI inTesla and Founding team of OpenAI) praises Tinker as a game-changer for AI researchers, simplifying LLM fine-tuning by retaining 90% creative control while slashing infrastructure complexity. He sees it enabling faster, smarter customization of smaller models for specific tasks, enhancing efficiency in production pipelines.

Implications for the AI Industry

Tinker’s emergence underscores a broader trend toward open AI development. By lowering entry barriers, it enables startups and academics to compete with tech behemoths, fostering diversity in AI applications. This could spur breakthroughs in personalized AI for sectors like education, where models tuned for specific curricula enhance learning outcomes, or in healthcare, for customized diagnostic tools.

However, it also raises ethical questions. Easier fine-tuning might amplify risks like model misuse for disinformation or biased outputs. Thinking Machines Lab’s emphasis on safety and transparency is commendable, but industry-wide standards will be crucial. Economically, the startup’s $12 billion valuation signals investor confidence in tools that support the AI supply chain, amid projections of the global AI market reaching trillions by 2030.

Challenges and Criticisms

Despite its promise, Tinker faces hurdles. Critics argue it doesn’t represent a “big-time blockbuster,” as it refines rather than reinvents fine-tuning processes. Competition from established platforms like Hugging Face or AWS could challenge adoption. Additionally, reliance on open-weight models limits its scope to non-proprietary systems, potentially alienating enterprise users seeking integration with closed APIs.

Scalability concerns loom, given the energy-intensive nature of AI training. As environmental impacts of data centers grow, Thinking Machines Lab must prioritize sustainable practices to maintain credibility.

Future Prospects

Looking ahead, Thinking Machines Lab hints at expanding Tinker’s capabilities, possibly incorporating multimodal models and additional products. Murati’s leadership positions the company to influence AI’s trajectory, emphasizing human-AI collaboration and ethical advancement.

In conclusion, Tinker represents a thoughtful step toward inclusive AI innovation. By empowering a wider audience to customize models, Thinking Machines Lab not only honors Murati’s legacy but also paves the way for a more equitable AI future. As the technology matures, its impact on research, industry, and society will undoubtedly unfold, reminding us that true progress lies in accessibility and shared knowledge.