What Thinking Machines Lab does

Thinking Machines Lab is a private, AI research and product company founded in February 2025 in San Francisco by former OpenAI CTO Mira Murati alongside a group of OpenAI alumni co-founders. The company's stated mission is to make AI systems more widely understood, customizable, and generally capable, with an explicit emphasis on human-AI collaboration rather than fully autonomous agents. Its product surface centers on two pillars: Tinker, a LoRA-based training and fine-tuning API launched in October 2025 that abstracts distributed compute for researchers running RL and supervised learning experiments across a broad lineup of open-source models (Qwen, GPT-OSS, DeepSeek, Moonshot Kimi, NVIDIA Nemotron); and TML-Interaction-Small, a proprietary 276B-parameter mixture-of-experts multimodal model (12B active) released as a research preview in May 2026 that natively processes audio, video, and text in 200ms time-aligned micro-turns and achieves state-of-the-art FD-bench v1.5 and turn-taking latency scores.

The underlying technology stack combines proprietary model architecture (encoder-free early fusion, dMel audio embedding, batch-invariant inference kernels) with significant published research contributions on the Connectionism blog covering LoRA fine-tuning methodology, on-policy distillation (demonstrating 9–30x compute cost reductions), and deterministic inference. The company has locked in frontier compute capacity through a multi-year, gigawatt-scale NVIDIA Vera Rubin deployment (targeting early 2027) and a multi-billion-dollar Google Cloud agreement on GB300-powered A4X Max VMs, with NVIDIA also making a significant equity investment.

Thinking Machines Lab's commercial model is primarily API-first and usage-based, with Tinker priced per million tokens across Prefill, Sample, and Train operations plus $0.10/GB-month storage. Early customers skew academic and research-oriented (UC Berkeley, Princeton, Stanford, Redwood Research), with enterprise revenue anchored by a $471.7M non-cancelable cloud-infrastructure contract with Boost Run and adjacent infrastructure services supplied to customers like RadixArk. Distribution runs through self-serve developer signup supplemented by direct enterprise agreements and cloud marketplace channels; revenue remains undisclosed and the company has been capital-intensive rather than profit-oriented since founding, with a $2B seed round at a $12B valuation in July 2025 underwriting ongoing buildout.

Name

Thinking Machines Lab

Legal name

Thinking Machines Lab, Inc.

Website

https://www.thinkingmachines.ai

Company type

Private

Founded year

2025

Operating status

Operating

Headcount range

11–50 employees

Short description

Thinking Machines Lab is a San Francisco-based AI research and product company building frontier multimodal interaction models (TML-Interaction-Small) and a fine-tuning API (Tinker), serving researchers, developers, and enterprises with customizable, real-time human-AI collaboration systems.

Ownership category

akta.pro rank

Product category

Foundation Models / AI Infrastructure

NAICS

Scientific Research and Development Services (5417), Computer Systems Design and Related Services (54151), Other Computer Related Services (541519)

SIC

Services-Computer Programming Services (7371)

akta.pro primary industry

Fine-Tuning, Adaptation & Custom Model Training (PEFT/LoRA/RLHF) (HDAAACAC)

akta.pro secondary industries

Model Development & Training Platforms (AutoML, Notebooks, Feature Stores) (HDAEANAB), LLMOps & Generative AI Platforms (Prompt/Agent Orchestration, RAG) (HDAEANAD), AI Compiler, Runtime & Kernel Optimization Software (CUDA/ROCm/XLA, graph compilers) (HDAAAAAI)

Keywords

• Multimodal AI models
• AI fine-tuning API
• Foundation model research
• AI infrastructure services
• Open-source AI training

Headquarters

HQ city

San Francisco

HQ country

United States

HQ region

North America

Offices2 records

Markets served

GTM type

B2B

Offering type

Software

Cost components

Infrastructure, Personnel, Technology or R&D, Marketing or Sales, Operations

Revenue model

1. Tinker API Consumption: Usage-based revenue from token consumption (Prefill, Sample, Train pricing per million tokens across multiple model families including Qwen, GPT-OSS, DeepSeek, Kimi, NVIDIA Nemotron). Storage charged at $0.10/GB-month. Pre-paid credits system with promotional grants available.
2. Enterprise Service Agreements: Pre-paid credits for Services (paid API/hosted model access); Enterprise Agreements possible for enterprise customers with proprietary terms; non-refundable fees with 1-year expiration on pre-paid credits
3. Multi-Billion-Dollar Cloud/Compute Contracts: Multi-year infrastructure deals with Google Cloud (single-digit billions) and NVIDIA (gigawatt-scale Vera Rubin deployment) for AI model training and deployment capacity

Pricing tiers

Go-to-market motion1 record

Distribution channels3 records

Marketing channels8 records

Core offering

Thinking Machines Lab builds and sells AI research products, primarily the Tinker fine-tuning API for researchers and developers (enabling LoRA-based training on open-source models like Qwen, GPT-OSS, DeepSeek, and NVIDIA Nemotron) and proprietary multimodal interaction models (TML-Interaction-Small) that natively handle real-time audio, video, and text collaboration with 200ms micro-turn latency. The company sells usage-based API access to its training platform and grants research preview access to its frontier interaction models.

Product overview

Thinking Machines Lab offers a platform-plus-modules product portfolio anchored on two main pillars: (1) Tinker, a managed fine-tuning API launched in October 2025 that lets researchers and developers apply LoRA-based supervised learning and RL training to a broad lineup of open-source models (Qwen, GPT-OSS, DeepSeek, Moonshot Kimi, NVIDIA Nemotron) without managing GPU infrastructure; and (2) TML-Interaction-Small, a proprietary 276B-parameter multimodal interaction model introduced as a research preview in May 2026 that natively handles full-duplex audio, video, and text interaction in 200ms micro-turns. The Connectionism research blog and News page round out the public-facing surface, publishing technical posts (LoRA Without Regret, On-Policy Distillation, Defeating Nondeterminism in LLM Inference), product announcements, and research grant programs (Interactivity Research Grants, Tinker Research and Teaching Grants) that bridge research outputs to the Tinker training platform and the company's frontier interaction models.

Differentiator

Problem solved

Functional benefit

Brands

• Tinker: A training API for researchers providing fine-tuning of open-source models with LoRA; handles scheduling, tuning, resource management, and infrastructure reliability. Includes a Tinker cookbook and a public X account @tinkerapi.
• Connectionism
• TML-Interaction-Small

Products and services

• Tinker A LoRA-based training and fine-tuning API for researchers and developers that exposes four functions (forward_backward, optim_step, sample, save_state) so users can run supervised learning and reinforcement learning experiments on a broad lineup of open-source models (Qwen, GPT-OSS, DeepSeek, Moonshot Kimi, NVIDIA Nemotron) without managing distributed compute infrastructure. Pricing is per million tokens with separate Prefill/Sample/Train rates and storage at $0.10/GB-month.
• TML-Interaction-Small Thinking Machines Lab's proprietary multimodal interaction model (276B parameter MoE with 12B active) that natively handles real-time full-duplex audio, video, and text interaction using 200ms time-aligned micro-turns with encoder-free early fusion. Achieves 0.40s turn-taking latency and 77.8 on FD-bench v1.5, outperforming GPT-realtime-2.0 and Gemini-3.1-flash-live. Currently available as a limited research preview.
• Connectionism Thinking Machines Lab's research blog named after the 1980s AI subfield. Publishes technical blog posts, papers, and code on topics including defeating nondeterminism in LLM inference, LoRA fine-tuning, on-policy distillation, and interaction models.

Named customers7 records

Ideal customer profiles3 records

Technology focussed Yes

API detail

Has API

Yes

API docs

API detail

Core technology

AI maturity

App detail

Integration9 records

AI capability13 records

Feature9 records

Partnerships

Nine partnerships are on record, tiered flagship and core.

• Google CloudflagshipStrategic or Co-development Partner · 22 April 2026Multi-billion dollar agreement giving Thinking Machines Lab access to Google Cloud's A4X Max virtual machines with NVIDIA GB300 GPUs as one of the first customers on Blackwell architecture. Includes Google Kubernetes Engine, Spanner, Cluster Director, Cloud Storage, and Anywhere Cache. Early testing showed 2X improvement in training and serving speeds.
• Jupiter Network (Google Cloud internal)coreTechnology or Integration · 22 April 2026Google Cloud's Jupiter network for weight transfers in reinforcement learning workloads, providing enhanced training performance through network optimization for AI workloads.
• Black Forest Labs, Cursor, LangChain, Mistral AI, Perplexity, Reflection AI, SarvamcoreStrategic or Co-development Partner · 20 March 2026Co-members of the NVIDIA Nemotron Coalition advancing open AI through shared research, expertise, data, and compute on NVIDIA DGX Cloud infrastructure. First joint project is a base model co-developed with Mistral AI.
• NVIDIA Nemotron CoalitioncoreStrategic or Co-development Partner · 16 March 2026Founding member of the NVIDIA Nemotron Coalition alongside Black Forest Labs, Cursor, LangChain, Mistral AI, Perplexity, Reflection AI, and Sarvam. Coalition will co-develop open frontier-level AI models trained on NVIDIA DGX Cloud infrastructure, with the first project being a base model co-developed with Mistral AI to underpin the Nemotron 4 family.
• Boost Run Inc. (BRUN)coreStrategic or Co-development Partner · 1 January 2026$471.7 million non-cancelable contract for AI cloud infrastructure services; reported as a meaningful contract win for the BRUN side while TML provides compute/infrastructure services.
• Google Cloud (Subcontractor)flagshipTechnology or IntegrationListed as authorized subcontractor providing global cloud computing infrastructure for the Tinker service, as documented in the DPA Annex 2.
• CloudflarecoreTechnology or IntegrationAuthorized subcontractor providing Content Delivery Network (CDN) services globally for the Tinker platform as documented in DPA Annex 2.
• WorkOScoreTechnology or IntegrationAuthorized subcontractor providing authentication services in the USA for the Tinker platform as documented in DPA Annex 2.
• Qwen (Alibaba)coreTechnology or IntegrationTinker supports Qwen family of models for fine-tuning, including Qwen3.6-35B-A3B, Qwen3.6-27B, Qwen3.5-4B, Qwen3.5-9B, and several other variants. On-policy distillation research used Qwen3-8B-Base with Qwen3-32B as teacher.

Scale indicators13 records

Recent moves8 records

Expansion highlights6 records

Emerging players

• OpenPipe: Managed fine-tuning and RL API for production LLM agents. Closest direct product analog to Tinker at the developer-platform layer for fine-tuning open models.
• Replicate: Cloud platform for running and fine-tuning open-source ML models via API. Overlaps with Tinker's API-first, open-model fine-tuning positioning for developers.
• Reflection AI: Frontier open-weight model startup and Nemotron Coalition co-member. Competes with TML on open frontier model development and shares NVIDIA compute infrastructure.

Direct peers

• Cohere: Enterprise-focused foundation model company offering customizable LLMs and fine-tuning APIs. Competes with TML on enterprise customization, though Cohere emphasizes closed models versus TML's open-model orientation.
• OpenAI: Frontier AI lab building foundation models (GPT-4o, Realtime API) and developer APIs. Most direct competitor to TML's Interaction Model and Tinker fine-tuning API; TML was founded by former OpenAI executives and shares the same multimodal interaction and customization thesis.
• Anthropic: Frontier AI lab with Claude models and an emphasis on AI safety and customization. Competes with TML on foundation model quality, developer tooling, and safety positioning, and has hired some former TML founders.
• Mistral AI: European open-weights model developer and Nemotron Coalition co-member alongside TML. Competes directly on open-weight foundation models and fine-tuning services that overlap with Tinker's value proposition.
• xAI: Elon Musk's frontier AI lab (Grok) backed by massive compute. Competes for the same frontier-research talent pool that TML has lost employees to, and operates in the same multimodal/generative AI category.
• Together AI: Open-model cloud platform offering fine-tuning, inference, and training APIs over open-source LLMs. Most directly comparable product competitor to Tinker's fine-tuning-as-a-service offering.

Broad incumbents

• Google DeepMind: Alphabet's frontier research lab producing Gemini and Gemini Live, which set the benchmarks TML's Interaction Model competes against. DeepMind is also the parent of Google Cloud, TML's multi-billion-dollar infrastructure partner.

Market position

Strengths5 records

Weaknesses5 records

Competitive moat4 records

Key risks7 records

Key highlights7 records

Customer concentration

• Xx.com/thinkymachines
• GitHubgithub.com/thinking-machines-lab
• LinkedInlinkedin.com/company/thinkingmachinesai

Compliance4 records

Revenue estimate

Valuation estimate

Number of profiles

Profiles5 records

Funding overview

Funding rounds3 records

Investors16 records

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A1 record

Investments1 record

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Thinking Machines Lab