Thinking Machines’ first official product is here: meet Tinker, an API for distributed LLM fine-tuning

Thinking Machines, the AI startup founded earlier this year by former OpenAI CTO Mira Murati, has launched its first product: Tinker, a Python-based API designed to make large language model (LLM) fine-tuning both powerful and accessible.
Now in private beta, Tinker gives developers and researchers direct control over their training pipelines while offloading the heavy lifting of distributed compute and infrastructure management.
As Murati wrote in a post on the social network X: “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.”
Tinker’s launch is the first public milestone for Thinking Machines, which raised $2 billion earlier this year from a16z, NVIDIA, Accel, and others.
The company’s goal is to support more open and customizable AI development — a mission that appears to resonate with both independent researchers and institutions frustrated by the opaque tooling around today’s proprietary models.
A Developer-Centric Training API
Tinker is not another drag-and-drop interface or black-box tuning service. Instead, it offers a low-level but user-friendly API, giving researchers granular control over loss functions, training loops, and data workflows — all in standard Python code.
The actual training workloads run on Thinking Machines’ managed infrastructure, enabling fast distributed execution without any of the usual GPU orchestration headaches.
At its core, Tinker offers:

Python-native primitives like forward_backward and sample, enabling users to build custom fine-tuning or RL algorithms.

Support for both small and large open-weight models, including Mixture-of-Experts architectures like Qwen-235B-A22B.

Integration with LoRA-based tuning, allowing multiple training jobs to share compute pools, optimizing cost-efficiency.

An open-source companion library called the Tinker Cookbook, which includes implementations of post-training methods.

As University of Berkeley computer science PhD student Tyler Griggs wrote on X after testing the API, “Many RL fine-tuning services are enterprise-oriented and don’t let you replace training logic. With Tinker, you can ignore compute and just ‘tinker’ with the envs, algs, and data.”
Real-World Use Cases Across Institutions
Before its public debut, Tinker was already in use across several research labs. Early adopters include teams from, yes, Berkeley as well as Princeton, Stanford,, and Redwood Research, each applying the API to unique model training problems:

Princeton’s Goedel Team fine-tuned LLMs for formal theorem proving. Using Tinker and LoRA with just 20% of the data, they matched the performance of full-parameter SFT models like Goedel-Prover V2. Their model, trained on Tinker, reached 88.1% pass@32 on the MiniF2F benchmark and 90.4% with self-correction, beating out larger closed models.

Rotskoff Lab at Stanford used Tinker to train chemical reasoning models. With reinforcement learning on top of LLaMA 70B, accuracy on IUPAC-to-formula conversion jumped from 15% to 50%, a boost researchers described as previously out of reach without major infra support.

SkyRL at Berkeley ran custom multi-agent reinforcement learning loops involving async off-policy training and multi-turn tool use — made tractable thanks to Tinker’s flexibility.

Redwood Research used Tinker to RL-train Qwen3-32B on long-context AI control tasks. Researcher Eric Gan shared that without Tinker, he likely wouldn’t have pursued the project, noting that scaling multi-node training had always been a barrier.

These examples demonstrate Tinker’s versatility — it supports both classical supervised fine-tuning and highly experimental RL pipelines across vastly different domains.
Community Endorsements from the AI Research World
The Tinker announcement sparked immediate reactions from across the AI research community.
Former OpenAI co-founder and former Tesla AI head Andrej Karpathy (now head of AI-native school Eureka Labs) praised Tinker’s design tradeoffs, writing on X: “Compared to the more common and existing paradigm of ‘upload your data, we’ll post-train your LLM,’ this is, in my opinion, a more clever place to slice up the complexity of post-training.”
He added that Tinker lets users retain ~90% of algorithmic control while removing ~90% of infrastructure pain.
John Schulman, former co-founder of OpenAI and now chief scientist and a co-founder of Thinking Machines, described Tinker on X as “the infrastructure I’ve always wanted,” and included a quote attributed to late British philosopher and mathematician Alfred North Whitehead: “Civilization advances by extending the number of important operations which we can perform without thinking of them.”
Others noted how clean the API was to use and how smoothly it handled RL-specific scenarios like parallel inference and checkpoint sampling.
Philipp Moritz and Robert Nishihara, co-founders of AnyScale and creators of the widely used open source AI applications scaling framework Ray, highlighted the opportunity to combine Tinker with distributed compute frameworks for even greater scale.
Free to Start, Pay-As-You-Go Pricing Coming Soon
Tinker is currently available in private beta, with a waitlist sign-up open to developers and research teams. During the beta, use of the platform is free. A usage-based pricing model will be introduced in the coming weeks.
For organizations interested in deeper integration or dedicated support, the company invites inquiries through its website.
Background on Thinking Machines and OpenAI Exodus
Thinking Machines was founded by Mira Murati, who served as CTO of OpenAI until her departure in September 2024. Her exit followed a period of organizational instability at OpenAI and marked one of several high-profile researcher departures, especially on OpenAI’s superalignment team, which has since been disbanded.
Murati announced her new company’s vision in early 2025, emphasizing three pillars:

Helping people adapt AI systems to their specific needs

Building strong foundations for capable and safe AI

Fostering open science through public releases of models, code, and research

In July, Murati confirmed that the company had raised $2 billion, positioning Thinking Machines as one of the most well-funded independent AI startups. Investors cited the team’s experience in core breakthroughs like ChatGPT, PPO, TRPO, PyTorch, and the OpenAI Gym.
The company distinguishes itself by focusing on multimodal AI systems that collaborate with users through natural communication, rather than aiming for fully autonomous agents. Its infrastructure and research efforts aim to support high-quality, adaptable models while maintaining rigorous safety standards.
Since then, it has also published several research papers on open source techniques that anyone in the machine learning and AI community can use freely.
This emphasis on openness, infrastructure quality, and researcher support sets Thinking Machines apart — even as the open source AI market has become intensely competitive, with numerous companies fielding powerful models that rival the performance of well-capitalized U.S. labs like OpenAI, Anthropic, Google, Meta, and others.
As competition for developer mindshare heats up, Thinking Machines is signaling that it’s ready to meet demand with a product, technical clarity, and public documentation.