# Thinking Machines Lab

> Source: https://aiwiki.ai/wiki/thinking_machines_lab
> Updated: 2026-10-27
> Fact-checked: 2026-10-27
> Categories: AI Companies, Generative AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Thinking Machines Lab." aiwiki.ai, 27 Oct 2026. https://aiwiki.ai/wiki/thinking_machines_lab
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Thinking Machines Lab** is an American [artificial intelligence](https://aiwiki.ai/wiki/artificial_intelligence) research and product company, headquartered in San Francisco, California, that was founded in February 2025 by [Mira Murati](https://aiwiki.ai/wiki/mira_murati), the former chief technology officer of [OpenAI](https://aiwiki.ai/wiki/openai), and raised a record $2 billion seed round at a $12 billion valuation in July 2025.[1][50] It is organized as a public benefit corporation.[51][4] At launch it said it wanted to make AI systems "more widely understood, customizable and generally capable"; since July 2026 it has described its mission as building "AI that extends human will and judgment."[56][38] Its six co-founders were drawn mostly from OpenAI's senior technical ranks and include [John Schulman](https://aiwiki.ai/wiki/john_schulman), an OpenAI co-founder who serves as chief scientist.[52][27]

The company drew attention early for the seniority of its founding team and the size of its seed round, and later for how quickly that team dispersed. It launched the research blog [Connectionism](https://aiwiki.ai/wiki/connectionism) in September 2025, released the fine-tuning service [Tinker](https://aiwiki.ai/wiki/tinker) in October 2025, previewed a real-time "interaction model" called TML-Interaction-Small in May 2026, and in July 2026 released [Inkling](https://aiwiki.ai/wiki/inkling), a 975-billion-parameter open-weight model, followed by the smaller Inkling-Small.[9][5][22][31][32] Over the same period it lost four of its six co-founders: Andrew Tulloch left for Meta in October 2025, Barret Zoph and Luke Metz returned to OpenAI in January 2026, and Lilian Weng rejoined OpenAI in July 2026.[13][11][47][48] A planned follow-on round at about $50 billion did not close; in September 2026 The Information reported that the company was in talks to raise $5 billion to $6 billion at a pre-money valuation of at least $40 billion, with Accel in talks to lead and Nvidia expected to supply about half.[29][34][45]

## How was Thinking Machines Lab founded?

Murati joined OpenAI in 2018, became its chief technology officer in 2022, and served briefly as interim chief executive after Sam Altman's abrupt firing in November 2023.[49][50] She announced her departure in September 2024, saying she wanted to do her "own exploration."[50][26][28] Her exit was one of several senior departures from OpenAI that autumn: Barret Zoph, a vice president of research, left the same day, and chief research officer Bob McGrew left around the same time.[49][53] In January 2025 Wired reported that Jonathan Lachman, previously OpenAI's head of special projects, had left to join Murati's new venture.[53]

The company came out of stealth on 18 February 2025 with a blog post that named Murati as chief executive, Schulman as chief scientist, and Zoph as chief technology officer.[50] TechCrunch counted 29 employees listed in the post, drawn from OpenAI, Character AI, Google DeepMind and other labs, while Reuters described a team of about 30 researchers and engineers from OpenAI, Meta and Mistral, roughly two thirds of them former OpenAI employees.[50][49][51] The company describes itself as "an artificial intelligence research and product company" and lists "research and product co-design" among its principles; its staff had worked on ChatGPT, Character.ai, Mistral's open-weight models, PyTorch, OpenAI Gym, Fairseq and Segment Anything.[56] In April 2025 TechCrunch reported that McGrew and [Alec Radford](https://aiwiki.ai/wiki/alec_radford), lead author of OpenAI's original GPT paper, had been added to the company's website as advisers.[54]

The full list of co-founders was not made public until the seed round closed in July 2025, when the company confirmed to Wired that they were Murati, Schulman, Zoph, [Lilian Weng](https://aiwiki.ai/wiki/lilian_weng), [Andrew Tulloch](https://aiwiki.ai/wiki/andrew_tulloch) and Luke Metz.[52]

## Who are the founders and leaders?

John Schulman, an OpenAI co-founder who joined Thinking Machines after a brief stint at [Anthropic](https://aiwiki.ai/wiki/anthropic), was a key figure behind [ChatGPT](https://aiwiki.ai/wiki/chatgpt) and has served as chief scientist since the launch.[11][51] [Barret Zoph](https://aiwiki.ai/wiki/barret_zoph), who had led model post-training at OpenAI, was the founding chief technology officer until January 2026.[50][11] Lilian Weng, who had worked on safety and robotics at OpenAI, was a co-founder until July 2026.[52][47]

### Founding team

| Name | Prior role | Role at Thinking Machines Lab | Status |
| --- | --- | --- | --- |
| [Mira Murati](https://aiwiki.ai/wiki/mira_murati) | CTO, [OpenAI](https://aiwiki.ai/wiki/openai) | Co-founder and CEO | Active |
| [John Schulman](https://aiwiki.ai/wiki/john_schulman) | Co-founder, [OpenAI](https://aiwiki.ai/wiki/openai); briefly at [Anthropic](https://aiwiki.ai/wiki/anthropic) | Co-founder and chief scientist | Active (as of September 2026) |
| [Barret Zoph](https://aiwiki.ai/wiki/barret_zoph) | VP of research (post-training), [OpenAI](https://aiwiki.ai/wiki/openai) | Co-founder and CTO | Departed January 2026 (to OpenAI; later reported to have moved to Google[45]) |
| [Lilian Weng](https://aiwiki.ai/wiki/lilian_weng) | Led Safety Systems, [OpenAI](https://aiwiki.ai/wiki/openai) | Co-founder | Departed July 2026 (to OpenAI) |
| [Andrew Tulloch](https://aiwiki.ai/wiki/andrew_tulloch) | Researcher, [OpenAI](https://aiwiki.ai/wiki/openai); earlier at Meta (Facebook) | Co-founder | Departed October 2025 (to Meta) |
| Luke Metz | Technical staff (post-training), [OpenAI](https://aiwiki.ai/wiki/openai) | Co-founder | Departed January 2026 (to OpenAI) |
| [Soumith Chintala](https://aiwiki.ai/wiki/soumith_chintala) | Co-creator of [PyTorch](https://aiwiki.ai/wiki/pytorch), [Meta](https://aiwiki.ai/wiki/meta_ai) | Chief technology officer from January 2026 | Active |
| Devendra Chaplot | Founding team, [Mistral AI](https://aiwiki.ai/wiki/mistral_ai) | Founding team (research and product, including Tinker) | Departed March 2026 (to [xAI](https://aiwiki.ai/wiki/xai) and SpaceX) |
| Joshua Gross | Engineer, [OpenAI](https://aiwiki.ai/wiki/openai) and [Meta](https://aiwiki.ai/wiki/meta_ai) | Founding engineer (Tinker) | Departed March 2026 (to Meta) |
| [Bob McGrew](https://aiwiki.ai/wiki/bob_mcgrew) | Chief research officer, [OpenAI](https://aiwiki.ai/wiki/openai) | Adviser | Named April 2025 |
| [Alec Radford](https://aiwiki.ai/wiki/alec_radford) | Researcher, [OpenAI](https://aiwiki.ai/wiki/openai) | Adviser | Named April 2025 |

When Zoph left on 14 January 2026, Murati announced in the same post that Soumith Chintala, co-creator of the [PyTorch](https://aiwiki.ai/wiki/pytorch) deep learning framework, would become chief technology officer, calling him "a major contributor to our team."[11] Chintala had left Meta in late 2025 after about eleven years there.[63] As of September 2026, Schulman remained chief scientist and Chintala ran technical work.[45]

Headcount grew despite the departures. Business Insider reported that the company had about 130 employees in April 2026 and, in May, that headcount had more than quadrupled since launch to over 150 people; by July 2026 TechCrunch put it at roughly 200.[23][16][34] TechCrunch's review of LinkedIn profiles in April 2026 found the company hiring more researchers from Meta than from any other single employer, including Segment Anything co-author Piotr Dollár, multimodal researchers Weiyao Wang and Andrea Madotto, and Kenneth Li, alongside hires such as Neal Wu, a founding member of the coding startup Cognition.[63] Business Insider reported that the company's careers page listed software engineering salaries of $350,000 to $475,000 a year.[16]

## How much funding has Thinking Machines Lab raised?

Thinking Machines Lab closed its $2 billion seed round in July 2025 at a $12 billion valuation, a company spokesperson told TechCrunch; The Next Web later described the figure as post-money.[1][45] The round was led by [Andreessen Horowitz](https://aiwiki.ai/wiki/andreessen_horowitz), with participation from [Nvidia](https://aiwiki.ai/wiki/nvidia), [Accel](https://aiwiki.ai/wiki/accel), [ServiceNow](https://aiwiki.ai/wiki/servicenow), Cisco, [AMD](https://aiwiki.ai/wiki/amd) and Jane Street.[1] Reports in June 2025 had put the valuation at $10 billion; it rose in the final weeks before the close.[1] Crunchbase News, writing in June 2025, called it "by far the largest seed round in the Crunchbase dataset," noting that the next-largest US seed financings had been in the $200 million to $450 million range, and Wired described it as the largest seed round in history.[2][52] Maginative reported that some funds passed on the deal because Murati's pitch offered no information about products or financial plans.[55]

### Financing history

| Date | Round or event | Amount | Valuation | Investors and status |
| --- | --- | --- | --- | --- |
| July 2025 | Seed | $2 billion | $12 billion post-money | Led by [Andreessen Horowitz](https://aiwiki.ai/wiki/andreessen_horowitz); closed |
| November 2025 | Proposed new round | Not disclosed | About $50 billion sought (some sources said $55 billion to $60 billion) | Talks reported by Bloomberg; stalled by January 2026 and did not close |
| March 2026 | Strategic investment | Not disclosed | Not disclosed | [Nvidia](https://aiwiki.ai/wiki/nvidia), alongside a compute partnership |
| September 2026 | Proposed new round | $5 billion to $6 billion (first reported as about $1 billion) | At least $40 billion pre-money | [Accel](https://aiwiki.ai/wiki/accel) in talks to lead; Nvidia about $2.5 billion to $3 billion; reported talks |

The investor list mixed [venture capital](https://aiwiki.ai/wiki/venture_capital) firms with strategic corporate backers, among them both Nvidia and AMD.[1] The government of Albania, Murati's country of origin, also decided to invest $10 million through its Ministry of Finance, proposing an amendment to the country's 2025 state budget to do so, according to the Albanian public broadcaster RTSH.[65]

In March 2026 Nvidia made what the companies called a "significant investment" as part of a multi-year compute partnership; the amount was not disclosed.[39][41]

### Did the $50 billion funding round close?

No. In November 2025 Bloomberg News reported that the company was in early talks to raise a new round at a valuation of roughly $50 billion, with some sources saying it could rise to between $55 billion and $60 billion; the terms had not been finalized.[29] TechCrunch reported in July 2026 that the round had been "said to be coming together last November but had stalled by January," and that the company had declined to discuss its funding since.[34] The Next Web described those talks as having not closed.[45]

### 2026 funding talks

On 3 September 2026 The Information reported that the company was in talks to raise new funding at a valuation of at least $40 billion, below the $50 billion it had sought in late 2025. TechCrunch, which carried the first version of the report, said the company was raising $1 billion with existing backer Accel in talks to lead.[44] The Information then updated its story: according to The Next Web and GuruFocus, the revised report put the raise at $5 billion to $6 billion at a pre-money valuation of at least $40 billion, with Nvidia expected to supply roughly half, about $2.5 billion to $3 billion.[45][46] The Next Web noted that the details came from Andreessen Horowitz briefing its own investors, according to The Information reporter Amir Efrati, rather than from a company announcement, and calculated that a raise at the top of the range would put the post-money valuation nearer $46 billion.[45] Neither Accel nor the company responded to TechCrunch's request for comment.[44]

### Revenue

The company does not publish revenue figures. TechCrunch reported in September 2026 that its annual revenue run rate was above $100 million, according to a source with knowledge of its finances, and that Inkling earns money through usage-based compute fees charged on Tinker.[44] The Next Web reported that The Information's Stephanie Palazzolo put annualized revenue "in the hundreds of millions," and, citing RuntimeWire, noted that the figure was self-reported and not audited.[45] In July 2026 the company said it had brought its technology to market and shown revenue in about nine months, according to TechCrunch.[34]

## What is Thinking Machines Lab's research mission?

At its launch the company set out three goals: making AI more widely understood, more customizable, and more generally capable.[50][56] Its website says that "science is better when shared" and that it plans "to frequently publish technical blog posts, papers, and code."[56] Its stated principles also include an emphasis on human-AI collaboration over fully autonomous systems, multimodal capability, and infrastructure quality.[56]

In July 2026 the company published an essay, "The Future Worth Building Is Human," that restated its mission as building "AI that extends human will and judgment."[38] The essay names four directions: training strong models with an emphasis on multimodal interaction and customizability; building tools that let people customize models, including training model weights; developing interfaces that widen the channel between people and AI; and publishing research.[38] It argues that most AI today is "trained in a handful of places and then frozen," and that alignment should come from "an ecosystem of AIs raised in different places" rather than a single model.[38] TechCrunch described the essay as the backdrop for the Inkling release the following week.[34]

The company's published research has clustered around [post-training](https://aiwiki.ai/wiki/post-training) and [reinforcement learning](https://aiwiki.ai/wiki/reinforcement_learning) methods, reproducible inference, optimization, and natively [multimodal model](https://aiwiki.ai/wiki/multimodal_model) architectures for real-time interaction.[57][22] The interaction-model post argues that "interactivity should scale alongside intelligence" rather than being added through external scaffolding around [foundation models](https://aiwiki.ai/wiki/foundation_models).[22]

### Connectionism research blog

In September 2025 the company launched **Connectionism**, a research blog it describes as "shared science from the team."[57] The first post, "Defeating Nondeterminism in LLM Inference," was published on 10 September 2025 by Horace He in collaboration with others at the company.[9] It argued that the main reason LLM inference endpoints give different outputs for the same prompt, even at temperature zero, is that their kernels lack "batch invariance," so a request's result depends on how many other requests share its batch.[9] In one experiment, sampling 1,000 completions from Qwen3-235B-A22B-Instruct with [greedy decoding](https://aiwiki.ai/wiki/greedy_decoding) produced 80 distinct outputs; with the company's batch-invariant kernels, all 1,000 were identical.[9] The work was taken up by open-source inference engines: the [SGLang](https://aiwiki.ai/wiki/sglang) team built deterministic inference on the company's batch-invariant operators later that month.[10]

| Date | Post | Author(s) | Subject |
| --- | --- | --- | --- |
| 10 September 2025 | Defeating Nondeterminism in LLM Inference | Horace He and others | Batch-invariant kernels for reproducible inference |
| 26 September 2025 | Modular Manifolds | Jeremy Bernstein | Constraining weight matrices to manifolds; a manifold version of the [Muon optimizer](https://aiwiki.ai/wiki/muon_optimizer) |
| 29 September 2025 | LoRA Without Regret | John Schulman and others | Conditions under which [LoRA](https://aiwiki.ai/wiki/lora) matches full fine-tuning, including in reinforcement learning |
| 27 October 2025 | On-Policy Distillation | Kevin Lu and others | Training a student model on its own samples with a teacher model's supervision |
| 11 May 2026 | Interaction Models: A Scalable Approach to Human-AI Collaboration | Thinking Machines | Real-time multimodal interaction models |
| 10 July 2026 | The Future Worth Building Is Human | Thinking Machines | Mission essay |
| 31 July 2026 | A Safe Path to Open Weights | Thinking Machines | Release policy for open-weight models |

## What products has Thinking Machines Lab released?

For most of 2025 Thinking Machines Lab kept its product plans secret, and TechCrunch noted at the time of the seed round that the company had yet to reveal what it was working on.[1] The first commercial product, Tinker, was unveiled on 1 October 2025, about seven and a half months after the company's public launch.[5][6]

### Tinker

**Tinker** is a managed [fine-tuning](https://aiwiki.ai/wiki/fine_tuning) service that exposes a Python API for adapting open-weight [large language models](https://aiwiki.ai/wiki/large_language_model) to specific tasks.[5][6] Users write the training loop and choose the data and algorithm, while the service runs the job on the company's own clusters and handles scheduling, resource allocation and failure recovery.[5] It exposes low-level primitives such as `forward_backward` and `sample`, which can express both [supervised fine-tuning](https://aiwiki.ai/wiki/supervised_fine-tuning) and reinforcement learning, and uses LoRA so that many training runs can share the same pool of compute.[5] The company also released an open-source library, the Tinker Cookbook, with implementations of common post-training methods.[5] Chief scientist John Schulman described it on launch day as "the infrastructure I've always wanted," calling it "an abstraction layer that is the right one for post-training R&D."[30]

The service launched in private beta on 1 October 2025, free at first with usage-based pricing to follow.[5] It became generally available on 12 December 2025, when the company ended the waitlist and added Moonshot AI's one-trillion-parameter [Kimi K2](https://aiwiki.ai/wiki/kimi_k2) Thinking, an OpenAI-API-compatible interface for sampling from checkpoints while they are still training, and image input through two Qwen3-VL models.[58][8] Tinker launched with 14 text-only models from Qwen and Llama; by October 2026 the company had added more than 20 more from Qwen, Nemotron, Kimi, GLM, DeepSeek, GPT-OSS and its own Inkling, ranging from 4 billion to 1 trillion parameters.[59] Over its first year it also added enterprise controls such as single sign-on, role-based access control, an audit-log API and private VPC connectivity, began offering full-parameter fine-tuning as a beta, and opened a sampling API compatible with OpenAI's and Anthropic's interfaces in beta.[59] The company says customer content on Tinker belongs to the customer and is never used for its own training.[59]

Early adopters named at launch were research groups at Princeton, Stanford, Berkeley and [Redwood Research](https://aiwiki.ai/wiki/redwood_research), which used it to train mathematical theorem provers, fine-tune a model for chemistry reasoning, run custom asynchronous multi-agent reinforcement learning loops, and train Qwen3-32B on AI control tasks.[5][7] In late October 2025 the company announced Tinker grants: $250 in credits per student for academic classes and research grants starting at $5,000.[60] Published customer projects have included forecasting models trained by Mantic with reinforcement learning on about 10,000 binary questions, and financial-document triage and text-to-SQL models built with Bridgewater Associates' AIA Labs.[67][68][69] TechCrunch reported that the Bridgewater financial model scored 84.7% on financial reasoning tests at about one fourteenth the running cost of leading proprietary models, results that came from the two companies' own evaluation.[34] The company has said Bridgewater's text-to-SQL model beat Claude Fable 5 and GPT-5.6 Sol on accuracy at 15% of their cost.[59]

### Interaction models

On 11 May 2026 Thinking Machines Lab published a research preview of what it calls **interaction models**, along with its first in-house model, **TML-Interaction-Small**.[22][24] The model is a sparse mixture-of-experts system with 276 billion total parameters and 12 billion active parameters.[22][23] It processes audio, video and text in continuous 200-millisecond "micro-turns," interleaving input and output so that it can listen and speak at the same time, interject, or stay silent.[22][24] This full-duplex design, in which the model takes in input while producing output rather than waiting for a turn to end, distinguishes it from earlier turn-based [voice assistant](https://aiwiki.ai/wiki/voice_assistant)s that rely on separate components to detect when a user has finished speaking.[22][23]

The architecture splits work between a live interaction model that stays in constant exchange with the user and a background model that handles slower reasoning, tool use and longer tasks asynchronously; the interaction model passes it the full conversation and weaves results back in as they arrive.[22] Audio enters the model as dMel spectrograms and images as 40-by-40-pixel patches through a lightweight embedding layer, without separate pretrained encoders.[22][23] The company reported a turn-taking latency of 0.40 seconds on the FD-bench V1 audio benchmark and a score of 77.8 on FD-bench V1.5, against 46.8 for GPT-realtime-2.0 at minimal reasoning.[22][23]

In demonstrations the model counted exercise repetitions in video and answered questions as visual evidence appeared, tasks measured by the RepCount-A and ProactiveVideoQA benchmarks; VentureBeat suggested uses such as monitoring a manufacturing video feed and flagging safety violations as they happen.[22][23] At the time of the announcement the models were not available to the public or to enterprises; the company said it would "open a limited research preview to collect feedback" in the coming months, "with a wider release later this year," and that its larger pretrained models were then too slow to serve in this setting.[22][23] On 19 May 2026 it announced interactivity research grants of $100,000 each, plus $25,000 in Tinker credits, for outside research on real-time interaction.[61]

### Inkling

On 15 July 2026 the company released **Inkling**, a model it trained from scratch and published with full weights on [Hugging Face](https://aiwiki.ai/wiki/hugging_face) under the [Apache License 2.0](https://aiwiki.ai/wiki/apache_license).[31][33] Inkling is a [mixture-of-experts](https://aiwiki.ai/wiki/mixture_of_experts) transformer with 975 billion total parameters and 41 billion active per token, a context window of up to one million tokens, and native input of text, images and audio, with text output.[31][33] The company was explicit that Inkling "is not the strongest overall model available today, open or closed," and presented it instead as a broad base model to customize through Tinker.[31][34]

| Property | Value |
| --- | --- |
| Release date | 15 July 2026 [31] |
| Parameters | 975B total, 41B active [31][33] |
| Layers | 66, decoder-only [33] |
| Experts | 256 routed experts (6 active per token) plus 2 shared [31][33] |
| Context window | Up to 1M tokens (64K and 256K tiers on Tinker) [31] |
| Pretraining data | 45 trillion tokens of text, images, audio and video [31] |
| Inputs and outputs | Text, image and audio in; text out [33] |
| Training hardware | NVIDIA GB300 NVL72 systems [31] |
| License | Apache 2.0 [33] |

According to the company, the expert layout largely follows [DeepSeek-V3](https://aiwiki.ai/wiki/deepseek_v3), attention interleaves sliding-window and global layers at a 5:1 ratio, and a relative positional embedding replaces the more common rotary embedding.[31] Pretraining used the [Muon optimizer](https://aiwiki.ai/wiki/muon_optimizer) for large weight matrices and Adam for other parameters, with schedules drawn from the company's modular-manifolds research.[31] Post-training began with supervised fine-tuning on synthetic data generated by open-weight models including Moonshot AI's [Kimi K2.5](https://aiwiki.ai/wiki/kimi_k2_5), and most post-training compute then went to large-scale asynchronous reinforcement learning over more than 30 million rollouts.[31] TechCrunch reported that the company said its next model would use fully self-contained post-training.[34] Inkling has an adjustable "thinking effort" setting; the company said it matched [Nemotron 3](https://aiwiki.ai/wiki/nemotron_3) Ultra on Terminal Bench 2.1 using about a third as many tokens.[31][34]

At launch Inkling was available on Tinker at a 50% introductory discount, through hosted APIs at [Together AI](https://aiwiki.ai/wiki/together_ai), [Fireworks AI](https://aiwiki.ai/wiki/fireworks_ai), [Modal](https://aiwiki.ai/wiki/modal), [Databricks](https://aiwiki.ai/wiki/databricks) and [Baseten](https://aiwiki.ai/wiki/baseten), and for local inference through SGLang, [vLLM](https://aiwiki.ai/wiki/vllm), TokenSpeed, llama.cpp and Hugging Face transformers; an NVFP4 checkpoint was published for Nvidia Blackwell systems.[31][33]

[Artificial Analysis](https://aiwiki.ai/wiki/artificial_analysis) scored Inkling at 41 on its Intelligence Index, calling it the leading open-weights release from a US lab, three points ahead of Nemotron 3 Ultra.[35] The Decoder reported that it still trailed the best Chinese open-weight models.[36] On Artificial Analysis's AA-Omniscience knowledge benchmark it scored +2, below the leading open-weight models, with a 63% hallucination rate.[35]

### Inkling-Small

Inkling-Small, previewed alongside Inkling, was released with open weights on 30 July 2026.[32] It is a mixture-of-experts model with 276 billion total parameters and 12 billion active, with the same one-million-token context window and native audio and image input.[32] The company said it changed the pretraining data mix and recipe for the smaller model, post-trained an earlier checkpoint partly through on-policy distillation from Inkling, and then ran two more weeks of agentic coding reinforcement learning.[32] It reported that Inkling-Small scored 31.6% on Humanity's Last Exam against Inkling's 29.7% and above 80% on SWE-bench Verified with a bash-only harness, while Inkling kept an advantage in knowledge coverage and factuality.[32]

### Safety and open weights

The company says it trained Inkling to an internal specification of safe behavior and commissioned outside testers.[31] In its July 2026 post "A Safe Path to Open Weights," it described three kinds of testing before releasing the weights: internal evaluations across a broad taxonomy of harms; pre-deployment red-teaming by four outside organizations ([Scale AI](https://aiwiki.ai/wiki/scale_ai) for general misuse, Handshake AI for interactions with vulnerable users, [FAR.AI](https://aiwiki.ai/wiki/far_ai) for chemical, biological, radiological, nuclear and cyber capabilities, and [Apollo Research](https://aiwiki.ai/wiki/apollo_research) for loss-of-control behaviors); and adversarial fine-tuning to strip safeguards and test worst-case capability.[37] It concluded that neither model extends the dangerous-capability frontier and that releasing Inkling was "not likely to add material risk beyond what existing open-weight models already pose," while noting that this balance would change as its models approach the frontier.[37] In August 2026 it offered Tinker credits of up to $50,000 for safety research on open-weight models.[62]

### Product roadmap

| Product | Type | Launch | Status |
| --- | --- | --- | --- |
| [Connectionism](https://aiwiki.ai/wiki/connectionism) | Research blog | September 2025 | Ongoing |
| [Tinker](https://aiwiki.ai/wiki/tinker) | Fine-tuning API | October 2025 (beta); December 2025 (general availability) | Generally available |
| TML-Interaction-Small | Native multimodal interaction model | May 2026 | Research preview |
| [Inkling](https://aiwiki.ai/wiki/inkling) | Open-weight multimodal model (975B MoE) | July 2026 | Weights and API available |
| Inkling-Small | Open-weight multimodal model (276B MoE) | July 2026 | Weights and API available |

## Partnerships and infrastructure

Thinking Machines Lab has signed several large compute agreements with chip and cloud providers.

On 10 March 2026 the company and Nvidia announced a multi-year partnership to deploy at least one gigawatt of Nvidia's next-generation [Vera Rubin](https://aiwiki.ai/wiki/nvidia_vera_rubin) systems for frontier model training and its customization platforms, with deployment targeted for early 2027.[39][40] The partnership also covers designing training and serving systems for Nvidia architectures, and came with Nvidia's undisclosed investment.[39][41] It deepened a relationship that began with Nvidia's participation in the seed round.[1]

On 22 April 2026, at the Google Cloud Next conference in Las Vegas, [Google Cloud](https://aiwiki.ai/wiki/google_cloud) announced a new agreement expanding the company's use of its infrastructure.[42] Thinking Machines became one of the first Google Cloud customers to use [Nvidia GB300 NVL72](https://aiwiki.ai/wiki/nvidia_gb300_nvl72) systems through A4X Max virtual machines, and Google said the company saw training and serving speed double in early testing compared with prior-generation GPUs.[42] A source told TechCrunch the deal was worth single-digit billions of dollars; TechCrunch also reported that it was not exclusive.[20][21] TechCrunch had reported in July 2025 that the company had previously struck a deal with Google Cloud to power its models.[1]

On 20 October 2026 Capital Brief reported that the AI infrastructure company Firmus had signed a multi-year deal under which Thinking Machines would become an anchor customer for Firmus's AI-Cloud Region in Asia, training and hosting its next-generation models on Nvidia GB300 systems.[43]

The company also builds much of its own systems software. Its interaction-model work uses batch-invariant kernels, the technique described in its first Connectionism post, to keep training and sampling bitwise aligned at less than 5% end-to-end overhead.[22][9]

## Why have so many founders departed?

From the start, Thinking Machines Lab's senior team was a target for competitive recruiting, and by July 2026 only Murati and Schulman remained of the six co-founders.[48] Business Insider reported in May 2026 that about 13 members of the founding team had left, including three of its six co-founders, and that some offers to founding members had reached "well into nine figures" in cash and stock over several years.[16] The departures went in three directions: to Meta, back to OpenAI, and to [Elon Musk](https://aiwiki.ai/wiki/elon_musk)'s xAI.[14][16]

In August 2025 the Wall Street Journal reported that [Mark Zuckerberg](https://aiwiki.ai/wiki/mark_zuckerberg) had sought to acquire Thinking Machines Lab and, after Murati declined, Meta pursued more than a dozen of its roughly 50 employees, including Tulloch, who was reportedly offered a package that could have been worth up to $1.5 billion over at least six years; a Meta spokesperson called that description "inaccurate and ridiculous."[13][23] Tulloch initially turned the offer down.[4] In October 2025 he left to join [Meta Superintelligence Labs](https://aiwiki.ai/wiki/meta_superintelligence_labs), with a company spokesperson saying he "has decided to pursue a different path for personal reasons."[13] Axios, citing a source, reported that the package he accepted was smaller than the earlier offer.[15] By April 2026, Business Insider reported, Meta had hired seven of the company's founding members, including Mark Jen and Yinghai Lu; founding engineer Joshua Gross, who had helped build Tinker, joined Meta Superintelligence Labs in March 2026.[23][14]

On 14 January 2026 Murati posted that the company had "parted ways with Barret Zoph." Within the hour, OpenAI applications chief [Fidji Simo](https://aiwiki.ai/wiki/fidji_simo) announced that Zoph, Luke Metz and researcher Sam Schoenholz were returning to OpenAI, writing that the move "has been in the works for several weeks."[11][12][17] Wired reported that the split with Zoph was not amicable.[11] TechCrunch described the simultaneous loss of two co-founders, one of them the CTO, as a potentially meaningful setback.[11]

In March 2026 Devendra Chaplot, a member of the founding team who had helped build Tinker and had previously been on the founding team of Mistral AI, announced that he was joining xAI and SpaceX, writing that he would work on building [superintelligence](https://aiwiki.ai/wiki/superintelligence).[66][18][19]

In late July 2026 Lilian Weng stepped down, writing that she had been sick more often in the previous seven months than at any point in her life and that the stress and workload had gone "beyond what my health can sustain physically," according to Business Insider as cited by The Next Web.[47] OpenAI confirmed within days that she would return to lead a research team focused on recursive self-improvement, using AI to help build future AI models.[47][48] Murati replied publicly that she was glad Weng was "putting your health first."[47]

Hiring continued alongside the losses. Headcount grew from about 130 in April 2026 to roughly 200 in July, and the company recruited heavily from Meta.[23][34][63]

## Reception and analysis

The $2 billion seed round was read as evidence that investors would fund a pre-product lab at a scale without precedent, largely on the strength of its founding team.[2][44] Coverage at the time noted that the company had yet to reveal what it was building.[1][55]

The October 2025 launch of Tinker drew praise from researchers. VentureBeat quoted [Andrej Karpathy](https://aiwiki.ai/wiki/andrej_karpathy), who called it "a more clever place to slice up the complexity of post-training" than the common "upload your data, we'll post-train your LLM" approach, and Berkeley PhD student Tyler Griggs, who said that unlike enterprise-oriented reinforcement-learning fine-tuning services, Tinker let users replace the training logic.[7]

Coverage of the May 2026 interaction-model preview focused on its latency and benchmark results, with caveats. VentureBeat pointed out that the models were not yet available to the public or to enterprises.[23] Semafor's Rachyl Jones wrote that the company's challenge was building something differentiated enough to draw users from OpenAI and Anthropic, since larger labs could follow with similar voice offerings.[25]

Inkling was described as the strongest open-weight model from a US lab at its release, but also as trailing leading Chinese open-weight models.[35][36] TechCrunch noted that the company had not said how it planned to cover its compute costs and that, because the weights are public, revenue has to come from Tinker and hosting rather than the model itself.[34] Its disclosure that it used Kimi K2.5 to generate some early post-training data was also noted: TechCrunch set it against wider scrutiny of training on competitors' model outputs, and The Decoder pointed out that Kimi K2.5 is a Chinese model.[34][36]

The wave of founder departures prompted commentary about whether a startup could hold a senior research team against the compensation offered by Meta and OpenAI.[14][16] Of the September 2026 talks, The Next Web reported Efrati's view that the price was lower than the company wanted but high for so young a lab, especially as a multiple of revenue, and estimated the implied multiple at roughly 80 to 400 times revenue.[45]

## Corporate structure and governance

The company's legal name is Thinking Machines Lab, Inc.[70] Murati structured it as a public benefit corporation, a form also used by OpenAI and Anthropic, to signal a commitment to developing AI that is "both accessible and beneficial to the public," according to eWeek.[51][4]

Reports on the seed round said that its terms gave Murati board voting rights that outweigh those of all other directors combined.[55] The Next Web described the arrangement in September 2026 as giving her voting control over the board on major corporate decisions "whatever the other directors do."[45]

## Headquarters and operations

The company is headquartered in San Francisco.[51] In August 2026 Commercial Observer reported that it had leased the entire fourth floor of 79 Fifth Avenue in Manhattan, about 21,500 square feet, as its first New York office.[64]

## Key milestones

| Date | Event |
| --- | --- |
| 25 September 2024 | [Mira Murati](https://aiwiki.ai/wiki/mira_murati) announces departure from [OpenAI](https://aiwiki.ai/wiki/openai) |
| 18 February 2025 | Thinking Machines Lab publicly announces its founding |
| April 2025 | [Bob McGrew](https://aiwiki.ai/wiki/bob_mcgrew) and [Alec Radford](https://aiwiki.ai/wiki/alec_radford) reported as advisers |
| July 2025 | $2 billion seed round closes at a $12 billion valuation; six co-founders confirmed |
| 10 September 2025 | Launch of the [Connectionism](https://aiwiki.ai/wiki/connectionism) research blog with "Defeating Nondeterminism in LLM Inference" |
| 1 October 2025 | Tinker fine-tuning API enters private beta |
| October 2025 | Co-founder [Andrew Tulloch](https://aiwiki.ai/wiki/andrew_tulloch) leaves for Meta |
| November 2025 | Talks reported for a new round at about a $50 billion valuation (later stalled) |
| 12 December 2025 | Tinker becomes generally available |
| 14 January 2026 | [Barret Zoph](https://aiwiki.ai/wiki/barret_zoph), Luke Metz and Sam Schoenholz depart for [OpenAI](https://aiwiki.ai/wiki/openai); [Soumith Chintala](https://aiwiki.ai/wiki/soumith_chintala) named chief technology officer |
| 10 March 2026 | Gigawatt-scale partnership and investment from [Nvidia](https://aiwiki.ai/wiki/nvidia) announced |
| March 2026 | Devendra Chaplot leaves for xAI; founding engineer Joshua Gross joins Meta |
| 22 April 2026 | Expanded [Google Cloud](https://aiwiki.ai/wiki/google_cloud) agreement announced |
| 11 May 2026 | Research preview of interaction models and TML-Interaction-Small |
| 10 July 2026 | Mission essay "The Future Worth Building Is Human" published |
| 15 July 2026 | [Inkling](https://aiwiki.ai/wiki/inkling) released with open weights |
| Late July 2026 | Co-founder [Lilian Weng](https://aiwiki.ai/wiki/lilian_weng) steps down and returns to OpenAI |
| 30 July 2026 | Inkling-Small released with open weights |
| August 2026 | First New York office leased; safety research grants announced |
| 3 September 2026 | Talks reported for a new round at a pre-money valuation of at least $40 billion |
| 15 October 2026 | Company marks one year of Tinker |
| 20 October 2026 | Multi-year compute deal with Firmus reported |

## See also

- [Tinker (Thinking Machines Lab)](https://aiwiki.ai/wiki/tinker)
- [Inkling (Thinking Machines Lab)](https://aiwiki.ai/wiki/inkling)
- [Mira Murati](https://aiwiki.ai/wiki/mira_murati)
- [John Schulman](https://aiwiki.ai/wiki/john_schulman)
- [Connectionism](https://aiwiki.ai/wiki/connectionism)
- [Open weights](https://aiwiki.ai/wiki/open_weights)
- [Safe Superintelligence Inc](https://aiwiki.ai/wiki/safe_superintelligence)
- [List of neolabs](https://aiwiki.ai/wiki/list_of_neolabs)
- [Neolab](https://aiwiki.ai/wiki/neolab)

## References

1. TechCrunch, "Mira Murati's Thinking Machines Lab is worth $12B in seed round," https://techcrunch.com/2025/07/15/mira-muratis-thinking-machines-lab-is-worth-12b-in-seed-round/, 15 July 2025.
2. Crunchbase News, "Thinking Machines Lab's $2B Seed Round Is Biggest By A Long Shot," https://news.crunchbase.com/venture/biggest-seed-round-ai-thinking-machines-mira-murati/, 23 June 2025.
3. Wikipedia, "Thinking Machines Lab," https://en.wikipedia.org/wiki/Thinking_Machines_Lab, accessed October 2026.
4. Built In, "Inside Thinking Machines Lab, Mira Murati's New AI Startup," https://builtin.com/articles/what-is-thinking-machines-lab.
5. Thinking Machines Lab, "Announcing Tinker," https://thinkingmachines.ai/news/announcing-tinker/, 1 October 2025.
6. SiliconANGLE, "Thinking Machines launches Tinker language model fine-tuning service," 1 October 2025.
7. VentureBeat, "Thinking Machines' first official product is here: meet Tinker," https://venturebeat.com/ai/thinking-machines-first-official-product-is-here-meet-tinker-an-api-for.
8. SiliconANGLE, "Thinking Machines makes its Tinker AI fine-tuning service generally available," 12 December 2025.
9. Thinking Machines Lab, "Defeating Nondeterminism in LLM Inference," https://thinkingmachines.ai/blog/defeating-nondeterminism-in-llm-inference/, 10 September 2025.
10. LMSYS Blog, "Towards Deterministic Inference in SGLang and Reproducible RL Training," https://lmsys.org/blog/2025-09-22-sglang-deterministic/, 22 September 2025.
11. TechCrunch, "Mira Murati's startup, Thinking Machines Lab, is losing two of its co-founders to OpenAI," https://techcrunch.com/2026/01/14/mira-muratis-startup-thinking-machines-lab-is-losing-two-of-its-co-founders-to-openai/, 14 January 2026.
12. Neowin, "Thinking Machines Lab CTO Barret Zoph returns to OpenAI in surprise move."
13. TechCrunch, "Thinking Machines Lab co-founder Andrew Tulloch heads to Meta," https://techcrunch.com/2025/10/11/thinking-machines-lab-co-founder-andrew-tulloch-heads-to-meta/, 11 October 2025.
14. The Next Web, "Meta hires five Thinking Machines Lab founders including a reported $1.5 billion engineer," https://thenextweb.com/news/meta-thinking-machines-lab-talent-raid, 21 April 2026.
15. Axios, "Behind the scenes: Meta still on prowl for AI talent," https://www.axios.com/2025/10/13/behind-the-scenes-meta-still-on-prowl-for-ai-talent, 13 October 2025.
16. American Bazaar, "One-third of Thinking Machines Lab founding team exits amid fierce AI hiring battle," https://americanbazaaronline.com/2026/05/14/one-third-of-thinking-machines-lab-founding-team-exits-480782/, 14 May 2026.
17. The Information, "OpenAI Hires Three AI Researchers From Murati's Thinking Machines Lab," https://www.theinformation.com/briefings/muratis-thinking-machines-lab-removes-cto.
18. The Information, "Elon Musk's xAI Hires Senior Thinking Machines Lab Staffer," https://www.theinformation.com/briefings/elon-musks-xai-hires-senior-thinking-machines-lab-staffer.
19. FinTech Weekly, "xAI's Superintelligence Push Just Got a Co-Founder of Mistral AI and a Member of Mira Murati's Founding Team," https://www.fintechweekly.com/news/xai-superintelligence-devendra-chaplot-mistral-thinking-machines-lab-grok-march-2026.
20. TechCrunch, "Exclusive: Google deepens Thinking Machines Lab ties with new multibillion-dollar deal," https://techcrunch.com/2026/04/22/exclusive-google-deepens-thinking-machines-lab-ties-with-new-multi-billion-dollar-deal/, 22 April 2026.
21. eWeek, "Mira Murati's Thinking Machines Lab Lands Billion-Dollar Google Cloud Deal."
22. Thinking Machines Lab, "Interaction Models: A Scalable Approach to Human-AI Collaboration," https://thinkingmachines.ai/blog/interaction-models/, 11 May 2026.
23. VentureBeat, "Thinking Machines shows off preview of near-realtime AI voice and video conversation with new 'interaction models'," https://venturebeat.com/technology/thinking-machines-shows-off-preview-of-near-realtime-ai-voice-and-video-conversation-with-new-interaction-models.
24. Unite.AI, "Thinking Machines Lab Ships First Model With 200ms Real-Time Interaction," https://www.unite.ai/thinking-machines-lab-ships-first-model-with-200ms-real-time-interaction/.
25. Semafor, "Mira Murati's Thinking Machines previews 'interaction models'," 13 May 2026.
26. Wikipedia, "Mira Murati," accessed June 2026.
27. Wikipedia, "John Schulman," accessed June 2026.
28. Fortune, "Meet Mira Murati, the 36-year-old tech prodigy who shot to fame at OpenAI," 3 October 2025.
29. Reuters via Yahoo Finance, "Mira Murati's Thinking Machines seeks $50 billion valuation in funding talks, Bloomberg News reports," https://finance.yahoo.com/news/mira-muratis-thinking-machines-seeks-212328387.html, 13 November 2025.
30. John Schulman (@johnschulman2), post on X, https://x.com/johnschulman2/status/1973450054238347314, 1 October 2025.
31. Thinking Machines Lab, "Inkling: Our Open-Weights Model," https://thinkingmachines.ai/news/introducing-inkling/, 15 July 2026.
32. Thinking Machines Lab, "Introducing Inkling-Small," https://thinkingmachines.ai/news/inkling-small/, 30 July 2026.
33. Hugging Face, "thinkingmachines/Inkling" model card, https://huggingface.co/thinkingmachines/Inkling, accessed October 2026.
34. TechCrunch, "Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling," https://techcrunch.com/2026/07/15/thinking-machines-amps-up-its-bet-against-one-size-fits-all-ai-with-its-first-open-model-inkling/, 15 July 2026.
35. Artificial Analysis, "Thinking Machines has released Inkling, the new leading U.S. open weights model," https://artificialanalysis.ai/articles/thinking-machines-has-released-inkling-the-new-leading-u-s-open-weights-model, 15 July 2026.
36. The Decoder, "Ex-OpenAI CTO Murati's Thinking Machines drops Inkling, a 975B parameter model that leads US labs but trails China," https://the-decoder.com/ex-openai-cto-muratis-thinking-machines-drops-inkling-a-975b-parameter-model-that-leads-us-labs-but-trails-china/, 16 July 2026.
37. Thinking Machines Lab, "A Safe Path to Open Weights," https://thinkingmachines.ai/blog/a-safe-path-to-open-weights/, 31 July 2026.
38. Thinking Machines Lab, "The Future Worth Building Is Human," https://thinkingmachines.ai/blog/the-future-worth-building-is-human/, 10 July 2026.
39. Thinking Machines Lab, "Thinking Machines Lab and NVIDIA Announce Long-Term Gigawatt-Scale Strategic Partnership," https://thinkingmachines.ai/news/nvidia-partnership/, 10 March 2026.
40. TechCrunch, "Thinking Machines Lab inks massive compute deal with Nvidia," https://techcrunch.com/2026/03/10/thinking-machines-lab-inks-massive-compute-deal-with-nvidia/, 10 March 2026.
41. CNBC, "Nvidia makes 'significant investment' in Mira Murati's Thinking Machines Lab," https://www.cnbc.com/2026/03/10/nvidia-mira-murati-thinking-machines-lab-ai.html, 10 March 2026.
42. Google Cloud (PR Newswire), "Thinking Machines Expands Use of Google Cloud AI Hypercomputer," https://www.googlecloudpresscorner.com/2026-04-22-Thinking-Machines-Expands-Use-of-Google-Cloud-AI-Hypercomputer, 22 April 2026.
43. Capital Brief, "Thinking Machines signs multi-year deal with Firmus," https://www.capitalbrief.com/briefing/thinking-machines-signs-multi-year-deal-with-firmus-81a979ea-90ee-4a4e-8560-436b0021286b/, 20 October 2026.
44. TechCrunch, "Accel reportedly in talks to lead $1B round for Thinking Machines at $40B valuation," https://techcrunch.com/2026/09/03/accel-reportedly-in-talks-to-lead-1b-round-for-thinking-machines-at-40b-valuation/, 3 September 2026.
45. The Next Web, "Thinking Machines is in talks to raise $6bn, The Information reports," https://thenextweb.com/news/thinking-machines-lab-6bn-round-nvidia-accel, 6 September 2026.
46. GuruFocus via Yahoo Finance, "Nvidia Is Eyeing a $2.5 Billion Bet on Mira Murati's AI Startup," https://finance.yahoo.com/technology/ai/articles/nvidia-eyeing-2-5-billion-142916646.html, 4 September 2026.
47. The Next Web, "She left the AI race over her health. Two days later, OpenAI had her back.," https://thenextweb.com/news/lilian-weng-thinking-machines-openai-recursive-self-improvement, 29 July 2026.
48. American Bazaar, "Lilian Weng leaves Thinking Machines to rejoin OpenAI amid health struggles," https://americanbazaaronline.com/2026/07/30/weng-leaves-thinking-machines-to-rejoin-openai-485496/, 30 July 2026.
49. Reuters via Yahoo Finance, "Former OpenAI tech chief Murati launches AI startup, poaching staff from ChatGPT maker," https://finance.yahoo.com/news/former-openai-technology-chief-mira-183245614.html, 18 February 2025.
50. TechCrunch, "Thinking Machines Lab is ex-OpenAI CTO Mira Murati's new startup," https://techcrunch.com/2025/02/18/thinking-machines-lab-is-ex-openai-cto-mira-muratis-new-startup/, 18 February 2025.
51. eWeek, "Sam Altman's Former CTO Launches OpenAI Competitor, Thinking Machine Labs," https://www.eweek.com/news/mira-murati-ai-startup-thinking-machines-lab/, 18 February 2025.
52. Wired, "Thinking Machines Lab Raises a Record $2 Billion, Announces Cofounders," https://www.wired.com/story/thinking-machines-lab-mira-murati-funding/, 15 July 2025.
53. Wired, "Mira Murati's AI Startup Makes First Hires, Including Former OpenAI Executive," https://www.wired.com/story/mira-murati-startup-hire-staff/, 17 January 2025.
54. TechCrunch, "Mira Murati's AI startup gains prominent ex-OpenAI advisers," https://techcrunch.com/2025/04/08/mira-muratis-ai-startup-gains-prominent-ex-openai-advisers/, 8 April 2025.
55. Maginative, "Former OpenAI CTO Mira Murati Raises $2B for Thinking Machines at $10B valuation," https://www.maginative.com/article/former-openai-cto-mira-murati-raises-2b-for-thinking-machines-at-10b-valuation/, 21 June 2025.
56. Thinking Machines Lab, homepage, https://thinkingmachines.ai/, accessed October 2026.
57. Thinking Machines Lab, "Connectionism," https://thinkingmachines.ai/blog/, accessed October 2026.
58. Thinking Machines Lab, "Tinker: General Availability and Vision Input," https://thinkingmachines.ai/news/tinker-general-availability/, 12 December 2025.
59. Thinking Machines Lab, "One Year of Tinker: Turning Your Expertise into Custom Models," https://thinkingmachines.ai/news/one-year-of-tinker/, 15 October 2026.
60. Thinking Machines Lab, "Tinker: Announcing Research and Teaching Grants," https://thinkingmachines.ai/news/tinker-research-and-teaching-grants/, 29 October 2025.
61. Thinking Machines Lab, "Announcing Interactivity Research Grants," https://thinkingmachines.ai/news/interactivity-research-grants/, 19 May 2026.
62. Thinking Machines Lab, "Announcing Safety Research Grants," https://thinkingmachines.ai/news/safety-research-grants/, 24 August 2026.
63. TechCrunch, "Meta's loss is Thinking Machines' gain," https://techcrunch.com/2026/04/24/metas-loss-is-thinking-machines-gain/, 24 April 2026.
64. Commercial Observer, "AI Startup Thinking Machines Lab Takes 22K SF at 79 Fifth Avenue for First NYC Office," https://commercialobserver.com/2026/08/thinking-machines-lab-79-fifth-avenue-lease/, 26 August 2026.
65. RTSH English, "Albania Enters Global AI Market; Government Invests $10 Million in 'Machine Thinking Lab' Led by Albanian-American Mira Murati," https://rtsh.al/rti/en/albania-enters-global-ai-market-government-invests-10-million-in-machine-thinking-lab-led-by-albanian-american-mira-murati/, 1 May 2025.
66. American Bazaar, "IIT graduate Devendra Chaplot to help Musk build 'superintelligence'," https://americanbazaaronline.com/2026/03/16/iit-graduate-devendra-chaplot-to-help-musk-build-superintelligence-476880/, 16 March 2026.
67. Thinking Machines Lab, "Training LLMs to Predict World Events (Guest Post with Mantic)," https://thinkingmachines.ai/news/training-llms-to-predict-world-events/, 19 March 2026.
68. Thinking Machines Lab, "Learning to Replicate Expert Judgment in Financial Tasks," https://thinkingmachines.ai/news/learning-to-replicate-expert-judgment-in-financial-tasks/, 30 June 2026.
69. Thinking Machines Lab, "Putting Task Expertise into RL Achieves State-of-the-Art Performance on Text-to-SQL," https://thinkingmachines.ai/news/putting-task-expertise-into-rl/, 27 August 2026.
70. Thinking Machines Lab, "Privacy Notice," https://thinkingmachines.ai/legal/privacy/, accessed October 2026.