Thinking Machines Lab

Thinking Machines Lab is a San Francisco AI company founded in early 2025 by Mira Murati, former Chief Technology Officer of OpenAI, following her September 2024 departure. Publicly announced in February 2025, it assembled a leadership team drawn heavily from OpenAI research and post-training staff, and is frequently discussed alongside Safe Superintelligence Inc. as part of the wave of companies founded by former OpenAI personnel. The company describes its goal as building "customizable, understandable" AI systems. It released its first research preview, a class of "interaction models," in May 2026, and its first full model — Inkling, an open-weight 975-billion-parameter mixture-of-experts system — on July 15, 2026 (Source: thinkingmachines.ai).

Snapshot

Funding and valuation

Overview and founding

Mira Murati founded Thinking Machines Lab after her September 2024 resignation from OpenAI, where she had served as Chief Technology Officer and, during the November 2023 governance crisis, as briefly-interim CEO. Her departure followed an extended period of senior exits from OpenAI: Ilya Sutskever and Jan Leike left in May 2024, Greg Brockman went on leave, and others departed subsequently. Thinking Machines is read as one outcome of that movement of staff; the other prominent example is Safe Superintelligence Inc., founded by Sutskever weeks after his own departure.

The company was publicly announced in February 2025, with a launch message emphasizing "customizable, understandable" AI and "collaborative general intelligence." Its early leadership concentrated former OpenAI research and post-training personnel, with additional hires from Anthropic and Meta. In July 2025, it reported a roughly $2 billion seed round at an approximately $12 billion valuation, described in reporting as one of the largest seed rounds in venture-capital history, with subsequent rounds also reported.

Stated mission

Thinking Machines's public communications emphasize three themes:

• Customizable — models that can be tailored to specific users, organizations, and contexts rather than a single monolithic assistant.
• Understandable — systems whose behavior is legible to users, connected to the broader interpretability discourse.
• Human–AI collaboration — a framing centered on augmentation rather than autonomy, presented in contrast with the more AGI-oriented framings at OpenAI and SSI.

Through early 2026 the company had not publicly detailed a training or deployment strategy, a model family, or a responsible-scaling policy.

Products

Through early 2026, Thinking Machines had not publicly released a model, API, or consumer product. That changed on May 11, 2026, with the research-preview release of a class of "interaction models" (Source: thinkingmachines.ai).

The preview describes a 276B-parameter mixture-of-experts model (12B active) built around 200-millisecond micro-turns that interleaves audio, video, and text streams, allowing the model to interrupt, backchannel, and react to visual cues without an external dialogue harness. The architecture removes the convention of dialogue as discrete turns in favor of continuous streaming at sub-turn latency. The company reported that TML-Interaction-Small leads its published FD-bench v1.5 benchmark at 77.8, with 0.40-second turn-taking latency (Source: thinkingmachines.ai). The release positions the company on real-time-interaction models rather than the chatbot or agentic-tool axis of competing frontier labs. It is adjacent to the "EQ as industry frontier" dynamic described by Wong (The Atlantic, April 2026) on the AI mental health page, though oriented toward conversational responsiveness rather than emotional intelligence as such.

The company also offers Tinker, a fine-tuning API. On June 30, 2026, Bridgewater's AIA Labs and Thinking Machines reported that a Qwen3-235B model fine-tuned via Tinker on expert-labeled data reached 84.7% accuracy on financial information-filtering tasks, above every frontier model the teams tested (best: 78.2%), at a 13.8× lower inference cost per task (Source: thinkingmachines.ai). See Financial Services — AI Deployment.

On July 15, 2026, the company released Inkling, its first in-house model: an open-weight mixture-of-experts transformer with 975 billion total parameters (roughly 41 billion active per task) and a 1-million-token context window, pretrained from scratch on 45 trillion tokens of text, image, audio, and video on Nvidia GB300 NVL72 systems, with full weights published on Hugging Face and fine-tuning offered through Tinker (Source: thinkingmachines.ai). A 276-billion-parameter Inkling-Small preview accompanied the release (Source: axios.com). The company stated that Inkling is "not the strongest overall model available today, open or closed," positioning it for enterprise fine-tuning rather than frontier-capability competition, and disclosed that early post-training data was partly generated using other open-weight models, including Moonshot AI's Kimi K2.5 (Source: techcrunch.com).

Nathan Lambert named Inkling and Inkling Small, both at 26 on the Artificial Analysis Intelligence Index as of September 14, 2026, as the leading American open models, behind the top Chinese open models at 42 to 45, and noted that the top American models were released in June and July 2026 and are updated less often than their Chinese counterparts. He cited Tinker as an example of the fine-tuning-API layer of an open-model post-training economy (The current balance of power in open models (Nathan Lambert, September 2026)).

People

• Mira Murati (Founder, CEO): Ex-OpenAI CTO (2022–2024); briefly interim CEO during the November 2023 governance crisis; oversaw the ChatGPT, DALL-E, Codex, and GPT-4 launches while at OpenAI. On June 4, 2026, she made what reporters described as her first major public appearance since founding the company, presenting an interaction-model vision at Bloomberg Tech in San Francisco; the appearance was characterized as a deliberately cautious re-emergence and the company's most visible external signal since the May 11 research preview (Source: msn.com).
• John Schulman (reported): Co-founder of OpenAI; co-inventor of PPO; briefly at Anthropic after leaving OpenAI.
• Barret Zoph (reported): Ex-OpenAI research; co-led post-training.
• Lilian Weng (reported): Ex-OpenAI; led Safety Systems; public writer on alignment. Her return to OpenAI, to work on recursive self-improvement, became public on August 3, 2026, days after she said that being a co-founder was taking a toll on her health and that she wanted a more focused position. She is the fourth co-founder to leave the company within the year to that date (Source: axios.com).
• Jonathan Lachman (reported): Ex-OpenAI.
• Andrew Tulloch (reported): Ex-Meta and ex-OpenAI. He returned to Meta in 2025, having declined a package the Wall Street Journal reported could have reached $1.5 billion before joining a few months later for an undisclosed sum, and announced his departure from Meta on September 9, 2026 without naming a destination. The same report describes him as the third Thinking Machines co-founder to change jobs in recent days, after Barret Zoph moved to Google and Luke Metz to Meta the previous month; the report frames both as departures from OpenAI, which the roster above does not record (Source: wsj.com).

Leadership attributions are triangulated from 2025 press reporting; specific titles and current rosters may have shifted.

Safety approach

Thinking Machines has not issued a public safety framework comparable to Anthropic's RSP or OpenAI's Preparedness Framework. The concentration of ex-OpenAI post-training leadership (Schulman, Zoph, Weng), several of whom were central to OpenAI's alignment and safety functions, has drawn both expectation that the company's safety posture will reflect their prior work and concern that these are the same researchers who built the systems current safety research seeks to understand.

Policy positions

Public policy positions are minimal to date. The company has not appeared as a prominent signatory or objector in the major legislative processes covered elsewhere, including California SB 53, the EU AI Act, and the Seoul commitments.

Relationships

• related: The current balance of power in open models (Nathan Lambert, September 2026) — Inkling as the leading American open model on the Artificial Analysis index (September 2026)
• related: Inkling Model Card (Thinking Machines Lab, July 2026) — the company's model card for Inkling (July 15, 2026), with its pre-release safety evaluations and conclusion of no material uplift beyond the open-weight ecosystem
• instance-of: AI Race Dynamics (talent-flow dimension)
• related: OpenAI (origin institution for most of leadership), Anthropic, Safe Superintelligence Inc. (SSI) (parallel company founded by former OpenAI staff), Inkling (first model release), Open-Weight Frontier Models (Inkling released open-weight)
• depends-on: RLHF (Reinforcement Learning from Human Feedback), Mechanistic Interpretability (intellectual lineage via leadership)

Murati is also referenced on Sam Altman.

Confidence note: Medium. Much of the detail above is triangulated from 2025 press reporting rather than primary company disclosures. Leadership roster and product direction may evolve.