Last reviewed
May 17, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 ยท 3,050 words
Add missing citations, update stale details, or suggest a clearer explanation.
Thinking Machines Lab is an American artificial intelligence research company headquartered in San Francisco, California. The company was founded in February 2025 by [[Mira Murati]], the former chief technology officer of [[OpenAI]], together with a group of researchers drawn predominantly from OpenAI's senior technical ranks. Structured as a public benefit corporation, Thinking Machines Lab states that its mission is to make AI systems more widely understood, customizable, and broadly capable, with an explicit emphasis on collaborative research and open scientific communication.
In its first year, the company drew unusual attention for both the seniority of its founding team and the scale of its early financing. In July 2025 it closed a $2 billion seed round at a $12 billion post-money valuation, the largest seed round on record in [[Andreessen Horowitz|Andreessen Horowitz's]] dataset and one of the largest first financings in Silicon Valley history. The lab released its first commercial product, the fine-tuning service Tinker, in October 2025, launched the research blog Connectionism in September 2025, and unveiled its first in-house model, TML-Interaction-Small, in May 2026.
Thinking Machines Lab grew out of the wave of senior researcher departures that followed the November 2023 governance crisis at [[OpenAI]]. Murati served briefly as OpenAI's interim chief executive officer during that crisis before returning to her CTO role, and announced her departure in September 2024 to pursue what she described as her "own exploration." Within weeks, several other prominent OpenAI researchers signaled their own exits, and by early 2025 a cluster of them had coalesced around Murati's new venture.
The company was incorporated as a Delaware public benefit corporation in early 2025 and emerged publicly on 18 February 2025 with a brief announcement that it was "building the future of artificial intelligence." Press coverage at the time noted that roughly thirty researchers and engineers had been hired from competitors including [[OpenAI]], [[Meta AI]], and [[Mistral AI]], with about two thirds of the initial technical staff coming from OpenAI. The company explicitly framed itself as a research-led organization that would also ship products, in contrast to pure research labs that defer commercialization.
Murati holds a controlling vote on most board matters under the company's bylaws, and founding shareholders hold votes weighted one hundred times those of regular shareholders. Reporting on the structure noted that the arrangement was intended to insulate long-term research direction from short-term investor pressure.
The founding cohort drew heavily on senior post-training, alignment, and applied-research talent from [[OpenAI]], joined by infrastructure and systems specialists from [[Meta AI]] and academic institutions. [[John Schulman]], an [[OpenAI]] co-founder who led its reinforcement learning team and is widely described as a principal architect of [[ChatGPT]], joined as chief scientist. Barret Zoph, previously vice president of research for post-training at OpenAI, served as the founding chief technology officer until his departure in January 2026. Lilian Weng, a former vice president at OpenAI known for her research blog on deep learning, joined as a founding researcher.
| Name | Prior role | Role at Thinking Machines Lab | Status |
|---|---|---|---|
| Mira Murati | CTO, OpenAI | Founder and CEO | Active |
| John Schulman | Co-founder, OpenAI; post-training lead | Chief Scientist | Active |
| Barret Zoph | VP of Research, OpenAI | Chief Technology Officer | Departed January 2026 |
| Lilian Weng | VP of Safety Systems, OpenAI | Founding Researcher | Active |
| Andrew Tulloch | Distinguished engineer, Meta AI | Founding Researcher | Departed mid-2025 |
| Luke Metz | Research scientist, OpenAI | Founding Researcher | Departed January 2026 |
| Jonathan Lachman | Head of special projects, OpenAI | Head of Operations | Active |
| Bob McGrew | Chief Research Officer, OpenAI | Adviser | Active |
| Alec Radford | Lead researcher, OpenAI | Adviser | Active |
Following Barret Zoph's January 2026 departure, the company recruited Soumith Chintala, co-creator of the [[PyTorch]] deep learning framework, from Meta to serve as chief technology officer. The hire was widely interpreted as a signal that Thinking Machines Lab intended to deepen its emphasis on training infrastructure and systems-level research.
By the start of 2026 the company employed roughly one hundred people and had grown its headquarters footprint in San Francisco accordingly. Recruiting was reported to have benefited from compensation packages that included substantial equity grants priced at the company's seed valuation.
Thinking Machines Lab's financing was, in the words of one analyst at Crunchbase, an outlier event "by a long shot" relative to the previous distribution of seed rounds. The company began accepting capital privately in early 2025, with the round closing publicly in July 2025 at a $12 billion post-money valuation. The financing was led by [[Andreessen Horowitz]], with participation from a broad syndicate that included [[Nvidia]], Accel, ServiceNow, Cisco, AMD, Jane Street Capital, and Conviction Partners. Earlier reports in June had suggested a $10 billion valuation, which rose in the final days of the deal as additional investors sought allocations.
| Date | Round | Amount | Post-money valuation | Lead investor |
|---|---|---|---|---|
| July 2025 | Seed | $2 billion | $12 billion | Andreessen Horowitz |
| March 2026 | Strategic | Undisclosed | Not disclosed | Nvidia |
The seed round drew commentary for the unusual concentration of strategic corporate investors alongside venture capital. Cisco and ServiceNow's participation reflected enterprise distribution interests, while [[Nvidia]] and AMD's joint involvement was unusual given the competitive dynamics between the two chip designers. A small allocation to the government of Albania, reportedly $10 million and requiring a national budget amendment, drew political attention in that country given Murati's Albanian heritage.
In March 2026, [[Nvidia]] made a follow-on strategic investment in connection with a multi-year capacity agreement for one gigawatt of computing built on its Vera Rubin platform. The amount of the equity component was not disclosed.
From its founding, Thinking Machines Lab has framed its mission in three connected goals: making AI more capable, making it more customizable to user needs, and making it more scientifically understood. The company has stated that scientific progress is best pursued collaboratively and that it intends to publish a substantial share of its research findings.
In practice, the company's early research output has clustered around three areas: post-training and [[reinforcement learning]] methods, inference systems and reproducibility, and natively multimodal model architectures. Researchers have argued publicly that the next major capabilities gains will come less from scaling pre-training compute and more from improvements to the techniques used to align and deploy [[foundation model|foundation models]] in interactive settings.
In September 2025 the company launched a research blog titled Connectionism, named for the historical school of thought that frames intelligence as emerging from networks of simple connected units. The blog's stated purpose is to share "scientific insights" with the broader research community. The first post, "Defeating Nondeterminism in LLM Inference," was authored by systems engineer Horace He and identified batch invariance as the underlying cause of irreproducibility in large language model outputs at temperature zero. The post showed that running an identical prompt 1,000 times against a Qwen-3-8B model with greedy decoding produced eighty distinct outputs, and proposed kernel-level changes that recovered exact reproducibility. The work was subsequently incorporated into open-source inference engines including SGLang.
Subsequent Connectionism posts have covered topics ranging from kernel numerics and asynchronous reinforcement learning to consistency in retrieval-augmented systems. The blog has also published code and benchmarks alongside several of its longer technical posts.
For most of 2025 Thinking Machines Lab maintained an unusual degree of secrecy about its product plans, with Murati and other executives declining to confirm specifics in public appearances. The first commercial product, Tinker, was unveiled at the start of October 2025, almost eight months after the company's founding.
Tinker is a managed [[fine tuning]] service that exposes a Python API for adapting open-weight [[large language model|large language models]] to specific tasks. The service launched in private beta in October 2025 and entered general availability in December 2025. Tinker allows developers to express a fine-tuning workflow as a single-process Python script, which the service then distributes automatically across multiple GPUs. The product supports a range of open-weight base models and exposes both supervised fine-tuning and reinforcement learning workflows.
Early adopters included research groups at Princeton, Stanford, Berkeley, and Redwood Research, who used the service for applications such as training mathematical theorem provers, fine-tuning models for chemistry reasoning, and prototyping custom asynchronous reinforcement learning training loops. The general availability release added a sampling tool that lets developers probe a model during training with test prompts and analyze its responses.
In May 2026 Thinking Machines Lab released its first in-house [[foundation model|foundation model]], TML-Interaction-Small, and a research preview of what the company calls interaction models. The model is a sparse mixture-of-experts system with 276 billion total parameters and 12 billion active parameters per token. It processes audio, video, and text in continuous 200-millisecond "micro-turns," interleaving input ingestion and output generation so that the model can listen and speak simultaneously, interject, or remain silent at conversational timing. The company reported a measured user-perceived response latency of about 0.40 seconds.
The interaction-model architecture splits computation between a live interaction model that maintains real-time presence with the user and a background model that performs slower reasoning and tool use asynchronously, with full conversation context shared between the two. The lab argued in its accompanying blog post that prior real-time voice systems had bolted streaming behavior onto turn-based architectures, whereas a natively multimodal and time-aware architecture was needed for the use cases it considered most valuable.
In demonstrations the model performed real-time visual tasks including repetition counting and proactive video question answering. The lab presented use cases such as a model monitoring a manufacturing video feed and proactively flagging safety violations. As of May 2026 the system is available only through a limited research preview to selected partners, with broader availability planned later in the year.
| Product | Type | Launch | Status |
|---|---|---|---|
| Tinker | Fine-tuning API | October 2025 (beta); December 2025 (GA) | Generally available |
| Connectionism | Research blog | September 2025 | Ongoing |
| TML-Interaction-Small | Native multimodal interaction model | May 2026 | Research preview |
Thinking Machines Lab has signed several large compute-capacity agreements with cloud providers and chip designers, reflecting the scale of training capacity required for its planned models.
In March 2026 the company announced a multi-year strategic partnership with [[Nvidia]] covering approximately one gigawatt of computing capacity built on the Vera Rubin platform, along with an undisclosed equity investment. The arrangement deepened a commercial relationship that began with Nvidia's participation in the seed round.
In April 2026 Google Cloud announced a multi-billion dollar multi-year agreement at the Google Cloud Next conference in Las Vegas, expanding a relationship that had begun in late 2025. Under the deal Thinking Machines Lab uses Google's A4X Max virtual machines, becoming one of the first Google Cloud customers to deploy Nvidia GB300 NVL72 systems. The company reported a roughly two-fold improvement in training and serving throughput versus the prior generation of accelerators. The agreement is not exclusive, and Thinking Machines Lab continues to use multiple cloud providers.
The lab's compute footprint is notable for its hybrid character. It runs training workloads on Nvidia GPUs leased through hyperscalers while also collaborating on inference optimizations that target both GPUs and tensor-processing accelerators. The Connectionism blog post on deterministic inference helped seed an open-source ecosystem of batch-invariant kernels that the lab uses internally and contributes to externally.
From the outset, Thinking Machines Lab's senior team was a magnet for competitive recruiting. Several founding researchers departed during the lab's first eighteen months in moves that were widely covered in the technology press.
In mid-2025, founding researcher Andrew Tulloch left to join Meta Superintelligence Labs as part of a broader recruitment campaign by [[Meta AI]] that reportedly included compensation offers of unprecedented size for individual researchers. Press accounts described several Thinking Machines Lab researchers as having received Meta offers, although most did not accept.
In January 2026 Barret Zoph and Luke Metz, both founding members, departed simultaneously to return to [[OpenAI]]. The dual departure was reported as a significant blow to the founding cohort and prompted commentary on the difficulty of retaining senior post-training talent in a market with multiple well-funded competitors. Within weeks, the company announced the hiring of Soumith Chintala from [[Meta AI]] as the new chief technology officer.
Despite the departures, the lab's overall headcount continued to grow, reaching approximately one hundred employees by early 2026. The company reported that hiring remained strong, with the share of new hires from [[OpenAI]] declining as the company broadened its recruiting into systems, infrastructure, and product roles.
Industry analysts have repeatedly cited Thinking Machines Lab as the most prominent test case for the thesis that a small, research-led startup founded by experienced talent can compete with established frontier labs. The $2 billion seed round was widely interpreted as evidence that capital markets were willing to fund such ventures at unprecedented scale, although critics noted that the valuation implied very high expectations relative to the company's pre-product status at the time of the financing.
The October 2025 launch of Tinker was reviewed favorably by developer-oriented publications, which praised the ergonomics of the Python API and the breadth of supported base models. Some commentators observed that the product targeted a different segment of the market from the consumer-facing chatbots offered by competitors such as [[OpenAI]] and Anthropic, suggesting that Thinking Machines Lab might pursue a more developer-platform-oriented strategy.
The May 2026 unveiling of interaction models drew more polarized reactions. Supporters argued that the architecture represented a genuine departure from existing voice and multimodal systems and that the sub-half-second latency would be hard to match without architectural redesign. Skeptics noted that the model had been released only as a small-scale research preview and that the gap between demonstration and a deployable, broadly accessible product remained substantial.
The departures of Zoph and Metz prompted commentary about the durability of large founding teams in a hot labor market, with some commentators arguing that the gravitational pull of [[OpenAI]] for senior post-training researchers had not been fundamentally weakened by the formation of competitor labs.
Thinking Machines Lab is organized as a Delaware public benefit corporation. Its certificate of incorporation specifies a stated public benefit related to advancing the responsible development and broad accessibility of AI systems. The structure resembles that of several other AI labs that have adopted public benefit forms while remaining for-profit corporations.
The company's governance arrangements include a board of directors with a weighted voting structure that gives Murati and other founding shareholders disproportionate voting power on most matters. Public reporting has described the arrangement as comparable to founder-control structures used by some publicly traded technology companies, but applied at the seed stage.
The company's headquarters is in San Francisco, California, in office space described as expanding to accommodate growth in research and product staff. The company operates substantially as a remote-flexible organization, with concentrations of staff in San Francisco and the broader Bay Area but distributed researchers and engineers across the United States.
| Date | Event |
|---|---|
| 25 September 2024 | Mira Murati announces departure from OpenAI |
| 18 February 2025 | Thinking Machines Lab publicly announces founding |
| July 2025 | $2 billion seed round closes at $12 billion valuation |
| 10 September 2025 | Launch of Connectionism research blog with "Defeating Nondeterminism in LLM Inference" |
| 1 October 2025 | Tinker fine-tuning API enters private beta |
| 12 December 2025 | Tinker becomes generally available |
| 14 January 2026 | Barret Zoph and Luke Metz depart for OpenAI |
| Late January 2026 | Soumith Chintala joins as chief technology officer |
| March 2026 | Multi-year strategic partnership with Nvidia announced |
| 22 April 2026 | Google Cloud multi-billion dollar agreement announced |
| 11 May 2026 | Research preview of TML-Interaction-Small and interaction-model architecture |