Thinking Machines Lab

  Executive Briefing

Thinking Machines Lab (TML) is a San Francisco AI research and product company founded in February 2025 by Mira Murati, the former Chief Technology Officer of OpenAI, with a stated mission to build "collaborative general intelligence" — AI systems that work with humans through natural, real-time multimodal interaction rather than rigid turn-based exchanges.¹ The company launched with roughly 30 researchers, approximately 20 of whom came directly from OpenAI, and quickly became one of the most closely watched AI startups in the world on the strength of its founding team's credentials alone. As of June 2026, TML is led by Mira Murati as CEO, John Schulman as Chief Scientist, Lilian Weng as co-founder and researcher, and Soumith Chintala as CTO.

TML's founding cohort reads as a who's-who of modern AI research. Schulman co-founded OpenAI and helped architect ChatGPT. Weng served as OpenAI's VP of Applied Research with a focus on safety. Chintala is the co-creator of PyTorch — the deep-learning framework underpinning the majority of modern AI research — and joined TML's technical staff in November 2025 before being elevated to CTO. The company was incorporated as a public benefit corporation, with Murati holding supervoting stock (100× voting power per share) and a deciding vote on the board, a governance structure designed to preserve her founding vision against investor influence.²

TML's technical bet is distinctive: rather than competing head-on with frontier incumbents through pure parameter scaling, the lab is pursuing efficient post-training optimization, native real-time multimodal architectures, and open, customizable fine-tuning infrastructure. Its first product, Tinker (launched October 2025), is a cloud API for fine-tuning open-weight large language models. Its first model, TML-Interaction-Small (research preview, May 2026), is a 276-billion-parameter Mixture-of-Experts system designed from the ground up for real-time audio, video, and text interaction, outperforming GPT-realtime-2.0 and Gemini on published interaction benchmarks.³

The company's first year was punctuated by exceptional turbulence alongside exceptional ambition. A reported $2 billion seed round closed at a reported $12 billion post-money valuation, later talks at a rumored $50–60 billion valuation stalled and appear to have collapsed, a co-founder (Barret Zoph) was publicly fired at an all-hands meeting in January 2026 for alleged misconduct and immediately rehired by OpenAI, and several other founding team members departed in quick succession.⁴⁵ Despite this volatility, TML has stabilized, announced major infrastructure partnerships with NVIDIA and Google Cloud, and shipped two live products. Whether it can translate its technical credibility into commercial traction and a viable funding path before its $2 billion in cash is depleted remains the central open question for the lab.

  At a Glance

  Origins & Founding

Thinking Machines Lab traces its origin to Mira Murati's departure from OpenAI in September 2024, where she had served as CTO since 2022 and had briefly acted as interim CEO during the November 2023 board crisis. Murati stated she was leaving to pursue "time and space for her own exploration," and within weeks she began assembling what would become an unusually concentrated founding team from OpenAI's senior ranks.¹ The founding thesis — that AI should be more widely understood, customizable, and built for natural human collaboration rather than one-sided API queries — directly reflected her experience at the frontier and her view that knowledge of how frontier models are trained had become dangerously concentrated inside a small number of labs.²

Between October 2024 and January 2025, Murati recruited five co-founders, each of whom held senior positions at OpenAI: John Schulman, a co-founder of OpenAI who had architected the reinforcement-learning-from-human-feedback training that powered ChatGPT (and who had briefly moved to Anthropic before joining TML); Barret Zoph, OpenAI's VP of Research for Post-Training; Lilian Weng, OpenAI's VP of Applied Research with a safety focus; Andrew Tulloch, a senior staff engineer specializing in large-scale training infrastructure; and Luke Metz, a senior staff researcher focused on evaluation systems.¹² The company was publicly announced in February 2025 with approximately 30 employees, with additional talent drawn from Meta AI and Mistral AI alongside the OpenAI alumni.

TML was incorporated as a public benefit corporation — a structure that requires balancing shareholder returns against social and public interests — and Murati secured supervoting rights (100× ordinary shares) plus a deciding vote on the board of directors, giving her effective veto power over major decisions even as external investors were brought in.² Advisers Bob McGrew (former OpenAI Chief Research Officer) and Alec Radford (former OpenAI lead researcher and co-creator of the GPT series and CLIP) joined in advisory roles, further cementing the lab's OpenAI lineage.

  History & Timeline

    Feb 2025 – Jun 2025: Launch and recruitment

TML launched publicly in February 2025 with roughly 30 researchers and a founding mission centered on "collaborative general intelligence." The early months were spent recruiting additional research talent, establishing the San Francisco office, and beginning foundational model and infrastructure work. At this stage TML had no commercial product and no revenue; investors were backing the team's credentials.

    Jul 2025 – Sep 2025: $2 billion seed round and first product work

TML closed what was reported as a $2 billion seed round at a reported $12 billion post-money valuation in approximately July 2025 (some sources cite June 2025; July appears the more commonly cited official close date).⁶ The round was led by Andreessen Horowitz and included Nvidia, AMD Ventures, Cisco, ServiceNow, Accel, Jane Street Capital, and other investors across a reported 23-participant syndicate — among the largest seed rounds in AI history.²⁶ Unusually, the Albanian government committed $10 million, requiring a domestic budget amendment, drawing international attention.² In September 2025, TML published research on reducing inconsistency in AI model outputs.

In August 2025, Meta launched a targeted recruitment campaign directed at TML researchers; initial approaches were largely declined.²

    Oct 2025 – Jan 2026: Product launch, departures, and crisis

TML launched Tinker on 1 October 2025 — a cloud API platform for fine-tuning open-weight large language models, initially in private beta. That same month, co-founder Andrew Tulloch departed TML to join Meta following Meta's talent push, becoming the first co-founder to leave.²

In November 2025, Bloomberg reported that TML had entered talks to raise an additional ~$5 billion at a rumored valuation of $50–60 billion.⁷ The round did not close; prospective backers reportedly declined to support the elevated valuation without a more substantial commercial track record. Soumith Chintala — the co-creator of PyTorch — joined TML's technical staff in November 2025 after departing Meta.²

The most disruptive event of TML's short history occurred on 14 January 2026, when Murati fired co-founder and CTO Barret Zoph at a company all-hands meeting, citing "serious misconduct."⁴ Reports from Wired initially suggested information-sharing with a competitor; later Wall Street Journal reporting indicated the primary issue was an undisclosed romantic relationship between Zoph and a junior colleague, with Zoph allegedly having lied when confronted.⁴⁵ Zoph denied the misconduct characterization, stating he was fired after Murati learned he was negotiating with another company, and that TML never cited performance issues. Within hours of the all-hands, OpenAI announced it was rehiring Zoph alongside co-founders Luke Metz and founding researcher Sam Schoenholz; OpenAI COO Fidji Simo stated OpenAI "does not share Murati's concerns" about Zoph.⁴ Researchers Lia Guy and Ian O'Connell also departed to OpenAI in the days that followed, and the all-hands itself was described as having "gone sideways" with employees resigning on the spot.⁵

Murati appointed Soumith Chintala as CTO on 15 January 2026, describing him as "a brilliant and seasoned leader."⁴

    Mar 2026 – present: Infrastructure partnerships and first model

TML announced a multi-year strategic partnership with NVIDIA on 10 March 2026, committing to deploy at least 1 gigawatt of next-generation Vera Rubin AI accelerators, with deployment targeted for early 2027; NVIDIA also made an undisclosed strategic investment in TML.⁸ On 22 April 2026, TML announced a multi-billion-dollar non-exclusive agreement with Google Cloud providing access to AI Hypercomputer infrastructure with Nvidia GB300 chips for frontier model training, reinforcement learning, and the Tinker platform — TML's first cloud services deal.⁹

On 12 May 2026, TML released TML-Interaction-Small as a research preview — a 276-billion-parameter Mixture-of-Experts model trained from scratch for real-time multimodal interaction, achieving 0.40-second turn latency and outperforming GPT-realtime-2.0 and Gemini on published interaction benchmarks.³

  Mission, Philosophy & Research Agenda

TML's mission is to build "collaborative general intelligence" — AI that works alongside humans through natural interaction modalities including speech, vision, and text, in real time and bidirectionally, rather than in the rigid request-response structure of conventional large language model APIs.² The company articulates a specific philosophical concern: that "knowledge of how frontier AI systems are trained is concentrated within the top research labs, limiting both public discourse on AI and people's abilities to use AI effectively," and that this concentration is itself a risk to the field's long-term health.²

From this foundation flow several distinguishing positions. TML prioritizes customizability and efficiency over raw parameter scaling, betting that fine-tuning and post-training optimization can yield competitive capability at lower compute cost. It commits to open research sharing, publishing technical work and building products (Tinker) designed to give researchers outside frontier labs access to fine-tuning infrastructure. And it frames the design target not as a general reasoning engine but as a collaborative partner — a system humans can shape, personalize, and interact with naturally over time. The research agenda reflects this: post-training optimization, Low-Rank Adaptation (LoRA) for efficient fine-tuning, real-time multimodal architectures, reinforcement learning from human feedback at scale, and methods for reducing model output inconsistency.

  Models & Products

TML has shipped two live products and one research-preview model as of June 2026, following the adaptation matrix guidance to fold research tightly into products for a small, young lab.

Tinker (launched 1 October 2025) is a cloud API platform for fine-tuning open-weight large language models. It supports fine-tuning of models including Meta Llama, Alibaba Qwen, DeepSeek V3.1, Moonshot AI Kimi K2, and OpenAI gpt-oss variants, using both supervised fine-tuning and reinforcement learning methods. Tinker handles automated orchestration, fault tolerance, and multi-node scaling, with pricing structured as a credit system metering orchestration work rather than raw GPU time. Early customers include the Princeton Gödel team, the Stanford Rotskoff lab, the UC Berkeley SkyRL lab, and Redwood Research, using the platform for reasoning models, multi-agent reinforcement learning, and AI safety research including backdoor detection.² As of launch, Tinker operated as a private beta with a waitlist.

TML-Interaction-Small (research preview, 12 May 2026) is TML's first original model. It is a 276-billion-parameter Mixture-of-Experts architecture with 12 billion active parameters per forward pass, designed from scratch as a native real-time multimodal system — trained end-to-end on audio, video, and text rather than assembled from separate modality-specific modules.³ Key architectural choices include an encoder-free early fusion design, dMel audio embeddings, hMLP video patch encoding, and a flow-head audio decoder. The model processes interaction in 200-millisecond micro-turn chunks, enabling simultaneous speaking and listening, visual proactivity, concurrent tool calls, web browsing, UI generation, and background reasoning.

Benchmark results at research-preview release include:

The research preview is available to a limited set of researchers as of May 2026, with wider release planned for later in 2026.³

➡️ See the individual model card for full specifications and benchmark details.

  People

TML's leadership structure has been substantially reshuffled from its founding composition. The current core team comprises Mira Murati (Founder and CEO), who holds supervoting stock and a deciding board vote; John Schulman (Co-founder and Chief Scientist), one of the field's foremost researchers in reinforcement learning and the principal architect of RLHF training for ChatGPT; Lilian Weng (Co-founder and Researcher), whose OpenAI career centered on safety-oriented applied research and who demonstrated TML-Interaction-Small in the May 2026 research preview; and Soumith Chintala (CTO, appointed January 2026), the co-creator of PyTorch and a figure whose technical credibility spans the entire deep-learning field.²⁴

Advisers include Bob McGrew (former OpenAI Chief Research Officer) and Alec Radford (former OpenAI lead researcher and co-creator of the GPT series and CLIP).²

The departures are as notable as the arrivals. Of the original six co-founders, three have left the company:

Additional researchers who departed in the January 2026 wave include Sam Schoenholz, Lia Guy, and Ian O'Connell, all returning to or joining OpenAI.⁴ As of April 2026, TML is reported to have approximately 140–169 employees, though the company's talent volatility makes this figure fast-moving.²

  Funding, Ownership & Business

TML raised what was reported as a $2 billion seed round at a reported $12 billion post-money valuation, closing in approximately July 2025 — one of the largest seed rounds in AI history at the time.⁶ The round was led by Andreessen Horowitz with participation from Nvidia (strategic investment), AMD Ventures, Cisco, ServiceNow, Accel, Jane Street Capital, Alpen Capital, Ambush Capital, and 15 additional investors across a reported 23-participant syndicate.²⁶ Unusually, the Albanian government committed $10 million — requiring a domestic budget amendment — making it one of a very small number of national governments to participate in a US AI startup seed round at this stage.²

The round closed with no commercial product shipped and no revenue; investors were explicitly backing the founding team's credentials. TML's business model centers on credit-based pricing for the Tinker platform, metering orchestration work rather than raw GPU time, though the company is understood to be pre-revenue in any material commercial sense as of mid-2025.²

In November 2025, Bloomberg reported that TML had entered discussions to raise an additional round of approximately $5 billion at a rumored valuation of $50–60 billion.⁷ Those talks are reported to have collapsed by January 2026 as prospective investors declined to support the elevated valuation without a more established commercial track record. As of June 2026, TML remains at its $12 billion paper valuation with approximately $2 billion in cash and no follow-on round closed; the timeline for revenue generation and the eventual need for additional capital remain key uncertainties.⁵

  Partnerships & Ecosystem

TML has moved quickly to secure the compute access a frontier model lab requires, signing two major infrastructure deals in the first quarter of 2026. On 10 March 2026, TML announced a multi-year strategic partnership with NVIDIA to deploy at least 1 gigawatt of next-generation Vera Rubin AI accelerator systems, with initial deployment targeted for early 2027; NVIDIA also made an undisclosed strategic investment in TML as part of the agreement.⁸ On 22 April 2026, TML announced a multi-billion-dollar non-exclusive agreement with Google Cloud providing access to AI Hypercomputer infrastructure equipped with Nvidia GB300 chips — claimed to deliver a 2× improvement in training and serving speed versus the prior generation — covering frontier model training, reinforcement learning workloads, and the Tinker platform.⁹ The Google Cloud deal is TML's first cloud services agreement.

On the research and early-adopter side, the Tinker platform has attracted a set of academic and safety-focused research customers: the Princeton Gödel team, the Stanford Rotskoff lab, the UC Berkeley SkyRL lab, and Redwood Research are among the early users, applying Tinker to reasoning model research, multi-agent reinforcement learning, and AI safety work including backdoor detection.²

  Compute & Infrastructure

Compute access is TML's most acute strategic constraint as it works toward frontier-scale model training. The NVIDIA Vera Rubin partnership (announced March 2026) represents TML's primary long-term bet on dedicated accelerator capacity: at least 1 gigawatt of Vera Rubin systems, with deployment beginning in early 2027, covers both frontier model training and the customizable AI platform work underlying Tinker.⁸ The Google Cloud agreement (April 2026) fills the near-term gap, providing GB300-based AI Hypercomputer access for current training runs, reinforcement learning, and inference.⁹ NVIDIA's separate strategic investment in TML aligns the chip supplier's incentives with TML's success in deploying and showcasing next-generation accelerators.

TML's architectural choices reflect its compute strategy: TML-Interaction-Small's Mixture-of-Experts design (276B parameters, 12B active) is explicitly engineered for inference efficiency, and the lab has published GPU kernel redesign work achieving batch-invariant inference with fewer than 5% performance overhead.² The 1-gigawatt Vera Rubin deployment, if realized on schedule in early 2027, would mark a step-change in TML's ability to train future frontier models at scale.

  Notable Events & Controversies

The January 2026 leadership crisis is the most significant single event in TML's brief history. On 14 January 2026, Murati fired co-founder and CTO Barret Zoph at a company all-hands meeting, publicly citing "serious misconduct."⁴ The exact nature of the alleged misconduct remains disputed. Wired initially reported it involved sharing confidential information with a competitor; the Wall Street Journal's subsequent reporting focused on an undisclosed romantic relationship between Zoph and a junior colleague — alleged to have begun during his time at OpenAI — and the claim that Zoph lied when confronted about it.⁴⁵ Zoph denied the misconduct characterization and told the WSJ he was fired after Murati learned he was negotiating with another company, with TML never raising performance concerns.

Within hours of the all-hands, OpenAI publicly announced it was rehiring Zoph alongside co-founders Luke Metz and founding researcher Sam Schoenholz. OpenAI COO Fidji Simo stated the company "does not share Murati's concerns" about Zoph — an extraordinary public rebuke of a rival lab's leadership decision.⁴ The all-hands meeting itself "went sideways," with multiple employees resigning on the spot.⁵

Fortune's reporting identified deeper structural tensions beneath the crisis: compensation disparity relative to established labs offering significantly richer packages to recruit talent; compute resource constraints versus what researchers had access to at incumbents; and internal frustration over the pace of product development and the lack of a clear business model.⁵ The collapse of the $50–60 billion valuation fundraise in the weeks preceding the crisis added public pressure, with commentators describing the original $12 billion seed round as "expensive FOMO" funding directed at "people, not businesses."⁵

The Albanian government investment ($10 million in the $2 billion seed round, requiring a domestic budget amendment) attracted separate international attention as an unusual instance of a small sovereign government's participation in a US AI startup at the seed stage.²

  Outlook & Roadmap

As of June 2026, Thinking Machines Lab has stabilized around a core leadership team of Murati, Schulman, Weng, and Chintala, with an estimated 140–169 employees and two live products: the Tinker fine-tuning platform and the TML-Interaction-Small research preview. The near-term roadmap centers on three tracks.

First, TML-Interaction-Small's release trajectory: the model is in limited research preview as of May 2026, with a broader release to researchers planned in the "coming months" and a wider public release described as coming "later in 2026"; no firm dates have been announced.³ Second, infrastructure scaling: the 1-gigawatt NVIDIA Vera Rubin deployment targeted for early 2027 represents TML's path to frontier-scale training capacity, and the Google Cloud agreement provides the bridge compute in the interim.⁸⁹ Third, Tinker's commercial expansion: the platform remains in private beta as of this writing; converting the waitlist into paying customers and demonstrating measurable revenue will be critical to validating TML's business model and supporting future fundraising.

On the financing side, the stalled $50–60 billion valuation round means TML will need to demonstrate meaningful commercial traction before a follow-on raise at an elevated valuation becomes viable. The $2 billion seed provides substantial runway, but the scale of compute investment required for frontier model development means the clock is running. Whether TML's technical differentiation — native real-time multimodal interaction and efficient fine-tuning infrastructure — translates into durable commercial advantage before that runway is exhausted is the defining strategic question the lab faces in 2026 and 2027.⁵⁷

---

  References

1. Thinking Machines Lab — Wikipedia
2. Thinking Machines Lab — Contrary Research
3. Interaction Models: A Scalable Approach to Human-AI Collaboration — thinkingmachines.ai
4. Mira Murati's startup Thinking Machines Lab is losing two of its co-founders to OpenAI — TechCrunch, Jan 2026
5. Mira Murati Thinking Machines staff defections — Fortune, Jan 2026
6. Thinking Machines Lab raises $2 billion — The Information
7. Murati's Thinking Machines in funding talks at $50 billion value — Bloomberg, Nov 2025
8. NVIDIA partnership announcement — thinkingmachines.ai
9. Google deepens Thinking Machines Lab ties with multi-billion dollar deal — TechCrunch, Apr 2026
10. Barret Zoph fired from Thinking Machines — Benzinga, Jan 2026
11. TML-Interaction-Small launch — MarkTechPost, May 2026
12. Mira Murati company financial breakdown — StartupHub, Jun 2026
13. What is Thinking Machines Lab — BuiltIn
14. Thinking Machines Lab Timeline — Brendon Beebe Substack

  References

1. TML founding and Murati's departure from OpenAI — Wikipedia; BuiltIn. ↩ ↩² ↩³

2. Company history, mission, structure, and investor details — Contrary Research; Wikipedia. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹² ↩¹³ ↩¹⁴ ↩¹⁵ ↩¹⁶ ↩¹⁷ ↩¹⁸ ↩¹⁹ ↩²⁰ ↩²¹

3. TML-Interaction-Small architecture and benchmarks — thinkingmachines.ai/blog/interaction-models/; MarkTechPost. ↩ ↩² ↩³ ↩⁴ ↩⁵

4. Barret Zoph firing and OpenAI's response — TechCrunch, Jan 2026; Benzinga. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰

5. Structural tensions, Fortune reporting on staff defections — Fortune, Jan 2026. ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹

6. $2 billion seed round details — The Information; exact close date (June vs. July 2025) varies by source; July 2025 appears the more commonly cited official close date. ↩ ↩² ↩³ ↩⁴

7. $50–60 billion valuation funding talks — Bloomberg, Nov 2025; round reported as collapsed by January 2026 per multiple sources, though TML has made no formal announcement. ↩ ↩² ↩³

8. NVIDIA Vera Rubin partnership — thinkingmachines.ai/news/nvidia-partnership/. NVIDIA's investment amount in TML is described as a "strategic investment" but the dollar figure is not publicly disclosed. ↩ ↩² ↩³ ↩⁴

9. Google Cloud multi-billion-dollar deal — TechCrunch, Apr 2026. ↩ ↩² ↩³ ↩⁴