Thinking Machines

Thinking Machines, formally Thinking Machines Lab, is an AI research and product company founded by Mira Murati and collaborators. Its work includes Tinker, a managed fine-tuning API; research on efficient adaptation and reproducible inference; real-time interaction models; and the open-weights Inkling model family. The company describes its goal as making AI more understandable, customizable and capable, with human–AI collaboration and research-informed product development as priorities.[1][3][6][11][12][14][15]

These products followed the company's public unveiling in February 2025, when it emphasized multimodal systems that people could adapt to their needs and expertise.[2][1]

Murati unveiled Thinking Machines Lab on February 18, 2025, after leaving her position as OpenAI's CTO the previous September. At launch, Murati was CEO, Barret Zoph was CTO and John Schulman was chief scientist; the company reported about 30 employees. A company-confirmed July founding roster named Murati, Schulman, Zoph, Lilian Weng, Andrew Tulloch and Luke Metz. Their OpenAI work spanned ChatGPT, safety and robotics, pretraining and reasoning, and post-training.[2][3]

Thinking Machines raised a $2 billion seed round led by Andreessen Horowitz, with participants including NVIDIA, Accel, Cisco and AMD. WIRED's July 15, 2025 account reported a $12 billion valuation; subsequent January reporting described the July round as closed. An earlier June 23 Crunchbase account had reported $10 billion. The two valuation accounts do not establish a pre-money versus post-money reconciliation.[3][25][4]

Tinker entered private beta on October 1, 2025. Developers choose their data, losses and training algorithms; Thinking Machines handles distributed execution, scheduling, resource allocation and failure recovery on its internal clusters. Launch models included Qwen-235B-A22B, a large mixture-of-experts model. Moving between supported model sizes required changing a model-name string in Python.[6][9]

The API's separate operations preserve control over how training proceeds: forward_backward computes losses and gradients, optim_step updates weights, and sample generates completions. A reinforcement-learning loop can sample responses, score them with rewards, train on those responses and repeat with updated weights. The open-source Tinker Cookbook supplies implementations for supervised fine-tuning, evaluation, distillation, preference learning, mathematical rewards and sandboxed code execution.[9][8]

At launch, Thinking Machines reported that Princeton's Goedel team had trained mathematical theorem provers, Stanford's Rotskoff Chemistry group had adapted a model for chemistry reasoning, Berkeley's SkyRL group had experimented with asynchronous reinforcement learning, and Redwood Research had trained Qwen3-32B on AI control tasks. Tinker became generally available on December 12, 2025, removing the waitlist and adding Kimi K2 Thinking, OpenAI API-compatible inference and Qwen3-VL vision inputs.[6][7][22]

Before Tinker's launch, two September 2025 research publications examined the adaptation method it uses and the numerical behavior of model serving. They supply technical context for sharing compute and running reproducible experiments.[6][10][11]

Tinker uses low-rank adaptation, or LoRA, to share compute across training runs. LoRA keeps original model weights fixed while training compact adapter matrices, reducing trainable state and allowing multiple adapted versions to share a base model. Adapter rank determines update capacity. In “LoRA Without Regret”, John Schulman and Thinking Machines reported that LoRA could match full fine-tuning when it had sufficient capacity and was applied across layers, especially MLP and mixture-of-experts layers. Larger datasets could exceed that capacity; attention-only adapters and some large-batch settings performed worse.[6][10]

The researchers recalled that early learning curves at different ranks looked so similar that they worried a bug was ignoring the rank parameter. Their analysis partly explained the similarity through the expected first update under α/r scaling, with α set to 32. Differences grew during longer training, and optimal learning rates then depended on rank. Their mathematical policy-gradient experiments matched full fine-tuning even at rank one; that result concerns those experiments rather than a universal Tinker configuration.[10]

Horace He and Thinking Machines' inference research explained how changing server load can change batch sizes, numerical reduction order and ultimately token selection—even at temperature zero. Their batch-invariant kernels addressed RMSNorm, matrix multiplication and attention. In a reported Qwen3-235B-A22B experiment, 1,000 requests produced 80 distinct completions without those kernels and identical completions with them. A separate Qwen3-8B serving test took 26 seconds with default vLLM, 55 seconds with unoptimized deterministic serving and 42 seconds after an attention-kernel improvement, illustrating the performance tradeoff.[11]

The founding roster changed substantially after launch. Andrew Tulloch left for Meta in October 2025. In January 2026, Zoph was dismissed and Luke Metz departed; OpenAI announced their return alongside researcher Sam Schoenholz. Murati appointed Soumith Chintala as CTO. Weng stepped down in July 2026, citing health reasons, and Murati publicly supported her decision; OpenAI confirmed Weng's return that month.[26][25][28][27]

The May 2026 interaction-model research preview introduced a model trained from scratch to process audio, video and text continuously. Its 200-millisecond micro-turns interleave incoming and outgoing streams, preserving silence, overlap and interruption in context. A foreground interaction model remains in conversation while a separate background model handles sustained reasoning and tool use asynchronously. The background model receives the conversation context, and its results stream back for integration into the ongoing exchange.[12]

The preview's TML-Interaction-Small had 276 billion total parameters and 12 billion active parameters. Some benchmark results included the background agent, measuring the combined system. Long-session context management, connectivity and the serving speed of larger models remained limitations. The 200-millisecond chunks specify processing granularity rather than guaranteed response latency; the announcement described limited access and wider release as future plans.[12]

Thinking Machines released Inkling on July 15, 2026, with downloadable weights and fine-tuning through Tinker. The mixture-of-experts transformer has 975 billion total parameters and 41 billion active per token, accepts text, images and audio, and generates text. Its architecture supports up to one million context tokens; release-time Tinker configurations offered 64,000 or 256,000. The company positioned it as a broad base for customization and a prospective background reasoning model for interaction systems.[14][16]

Inkling was pretrained on 45 trillion tokens spanning text, images, audio and video using NVIDIA GB300 NVL72 systems. Thinking Machines reported more than 30 million asynchronous reinforcement-learning rollouts across two long continuous runs. A company demonstration made customization concrete: Inkling wrote and ran a Tinker fine-tuning job to avoid the letter “e,” evaluated the result and staged a checkpoint for a supervisor to relaunch inside OpenCode. Another demonstration refined a multiplayer snake game through 40 rounds of feedback from GPT Codex.[14]

Inkling-Small has 276 billion total parameters and 12 billion active. Training began after Inkling, allowing changes to the pretraining data mix and recipe. Its earlier student preview checkpoint was partly post-trained through on-policy distillation with Inkling as teacher; the team then continued agentic coding reinforcement learning for two weeks. Both models offer controllable thinking effort to balance performance and computation.[15][14]

The company's released comparisons show task-specific tradeoffs: Inkling-Small scored 31.6% versus Inkling's 29.7% on text-only Humanity's Last Exam and 80.2% versus 77.6% on SWE-bench Verified, but scored lower on SimpleQA Verified and Tau 3 Banking. The declared evaluation settings were effort 0.99 and temperature 1.0, with a 256,000-token limit for coding trajectories; SWE-bench results used a bash-only harness.[15]

Downloaded Inkling weights require substantial hardware: its model card specifies at least 2 TB of aggregate GPU memory for BF16 or 600 GB for NVFP4, with configuration-specific GPU requirements. The card declares Apache 2.0 and links a separate Model Acceptable Use Policy. The company asserts conditions covering downloads, use and modified model materials, restrictions beyond illegal conduct, and responsibilities for downstream users.[16][17]

Tinker also supported collaborations using other organizations' base models. A June 2026 financial-task case preceded the Inkling releases; an August text-to-SQL case extended this work on task-specific data and rewards.[20][21][14]

In work published June 30, Bridgewater AIA Labs authors Sarah Su, Kevin Zhu, Emily Xiao, Rohan Alur and Daniel Kang collaborated with Thinking Machines to adapt Qwen3-235B on Tinker for six financial-document tasks. These included identifying relevant articles and central-bank documents, answering document-relevance questions and separating useful content from boilerplate. Initial vendor labels produced poor models; the team routed disagreements between model outputs and labels to investment experts for correction, then evaluated on held-out data.[20]

The recipe combined interleaved task batches, CISPO loss with asymmetric clipping and on-policy distillation. Every 20 steps, a checkpoint could replace the teacher only after reaching a new validation-accuracy high. The collaborators reported 84.7% average accuracy, versus 78.2% for the best tested frontier baseline, corresponding to 29.8% fewer errors, with 13.8-fold lower inference cost per task. Accuracy measured agreement with investor labels across the six tasks; it did not measure investment returns.[20]

The August 27 text-to-SQL collaboration credited UIUC researchers Yuxuan Zhu, Tengjun Jin and Yoojin Choi, and Bridgewater/UIUC collaborator Daniel Kang, working with Thinking Machines. They fine-tuned Moonshot's Kimi-K2.6 through Tinker to produce ReViSQL-K2.6. Expert correction produced BIRD-Platinum training data; a separate second correction pass over evaluation references, following Arcwise's cleanup of BIRD Mini-Dev, produced Arcwise-Plat-SQL.[21]

The collaborators addressed two ways a correct-looking result can teach the wrong skill. Averaging line-item prices can match averaging order totals when every order happens to contain one item. They used VeriEQL's bounded verification to penalize execution matches it refuted. Likewise, sodium < 5 could return the same rows as a supplied definition requiring sodium = 0; rule-based process rewards required the model to translate external knowledge into query constraints and audit its query against them.[21]

On the corrected Arcwise-Plat-SQL evaluation set, the collaborators reported 88.55% accuracy from verified-data training alone and 91.37% with both reward modifications under greedy decoding, at $0.035 per task. Generating 16 candidates at temperature one and selecting a query from the majority execution-result group reached 92.97%, at $0.56 per task. These measurements concern the corrected evaluation set.[21]

Infrastructure planning preceded the August training results. On March 10, 2026, Thinking Machines and NVIDIA announced a partnership committing to deploy at least one gigawatt of Vera Rubin systems, targeting early 2027. The partnership includes designing training and serving systems for NVIDIA architectures. NVIDIA also disclosed an investment without stating its amount. The gigawatt figure describes planned capacity.[18]

Thinking Machines announced interactivity research grants on May 19, 2026, advertising multiple grants of $100,000 each plus $25,000 in Tinker credits. On August 24, it announced safety research grants of up to $50,000 in Tinker credits for open-weight-model research. These are advertised grant amounts, rather than documented aggregate spending or paid awards.[23][24]

The company's direction essay described its mission as building AI that “extends human will and judgment.” It connected frontier models, weight customization, live multimodal interfaces and published research to that aim, arguing that people and organizations should continuously shape the systems they use.[13]

On September 3, 2026, TechCrunch reported an annual revenue run rate above $100 million, citing an unnamed source familiar with the company's finances. That is a reported annualized measure rather than audited annual revenue. September reporting also described further financing discussions; accounts differed on the proposed amount, and the discussions did not establish a closed round.[29][31]