Thinking Machines Lab began as former OpenAI CTO Mira Murati’s AI research and product startup, announced on February 18, 2025. Its original pitch centered on customizable, multimodal AI designed to work with people. By August 18, 2026, the company had launched Tinker, published research on real-time Interaction Models, released the open-weights Inkling model and announced a major NVIDIA compute partnership. The story is now less about a launch-day mission statement than whether those products can turn an unusually large early investment into useful, dependable tools.
What Thinking Machines Lab is—and what it set out to do
Thinking Machines Lab describes itself as an AI research and product company. Its February 2025 launch framed a broad goal: make AI systems more widely understood, customizable and generally capable, with an emphasis on human-AI collaboration rather than only autonomous systems. The company also pointed to multimodal interaction, frontier work in science and programming, open scientific work, and empirical approaches to AI safety. Its stated priorities are outlined on the company website.
As an Amazon Associate I earn from qualifying purchases.
That announcement established a direction, not a product specification. At launch, the company had not disclosed a specific product, detailed model architecture, public pricing or a release timetable, according to Axios’s February 2025 report. The distinction matters: the launch described what the founders wanted to build, while later releases and research provide more concrete evidence of how they are pursuing it.
Free tools Windows power users keep installed
One-click scans. No signup required.
Who founded it?
Murati joined OpenAI in 2018, became its chief technology officer in 2022 and briefly served as interim chief executive during the November 2023 leadership crisis. She was associated with OpenAI efforts including ChatGPT, DALL·E and Codex, and announced her departure in 2024 before launching Thinking Machines. She is the startup’s cofounder and CEO; “former OpenAI CTO” describes her previous role, not a company affiliation. TechCrunch covered her background and the launch.
#1 Best Overall
- EVOLUTION AMD RYZEN AI MAX+ 395 MINI PC - GMKtec EVO-X2 is the next evolution in AI mini PC Ryzen Strix Halo series. Thanks to AMD Simultaneous Multithreading (SMT) the core-count is effectively doubled, to 32 threads. Ryzen AI Max+ 395 has 64 MB of L3 cache and can boost up to 5.1 GHz, depending on the workload. The Ryzen AI Max+ 395 is currently rated as the "most powerful x86 APU" on the market for AI computing.
- AI NPU with XDNA 2 ARCHITECTURE - Powered by 16 “Zen 5” CPU cores, 50+ peak AI TOPS XDNA 2 NPU and a truly massive integrated GPU driven by 40 AMD RDNA 3.5 CUs, the Ryzen AI MAX+ 395 is a transformative upgrade and delivers a significant performance boost over the competition. The Ryzen AI Max+ 395 excels in consumer AI workloads like the llama.cpp-powered application: LM Studio. Shaping up to be the must-have app for client LLM workloads, LM Studio allows users to locally run the latest language model without any technical knowledge required and unleash their creativity and productivity.
- AMD RADEON 8090S iGPU GAMING PC - The AMD Radeon RX 8060S offers all 40 CUs with up to 2.9 GHz graphics clock and uses the new RDNA 3.5 architecture. The powerful iGPU is positioned between an RTX 4060 and 4070 laptop GPU and therefore enables gaming in FHD at maximum details in most demanding games. The 8060S can also utilize the full 128GB pool, which is perfect for running LLMs such as Deepseek 70B Q8, which runs comfortably on this machine.
- EIGHT CHANNEL LPDDR5X - LPDDR5X is a new ground breaking memory small form factor installed on-board. With blazing speeds up to to 8000MT/s, it runs 1.5x faster than the DDR5 SODIMMs; 90% better performance over DDR5 SODIMMs in video conferencing and photo editing; 30% better performance in productivity apps; 12% better performance in digital content workloads.
- QUAD SCREEN 8K DISPLAY SUPPORT - EVO-X2 AI Mini PC support 4-screen 4K/8K output via HDMI 2.1 (8K@60Hz), DisplayPort 1.4 (4K@60Hz), and dual USB 4 40Gbps Transfer speed (supporting PD3.0/DP1.4/DATA). Ideal for gaming, video editing, and multitasking, it provides expansive and crisp multi-display support.
The early research team
The launch brought together several prominent researchers, including John Schulman, an OpenAI cofounder and reinforcement-learning researcher named chief scientist; Barret Zoph, named CTO; and researchers Lilian Weng, Andrew Tulloch and Luke Metz. Early reporting put the team at about 30 people, a launch-era figure rather than a current headcount. Axios reported the early team size, while WIRED covered the founding team and financing.
Why multimodality and human-AI collaboration are central
“Multimodal” here means more than a chatbot that accepts an image. The company’s stated vision spans text, audio, video and visual context, and imagines systems that can follow conversation, respond to interruptions, use tools and generate interfaces. In principle, that could let a system use more of the context a person provides and interact in ways closer to ordinary collaboration. Multimodal input alone, however, does not establish better reasoning or performance.
The company’s May 2026 research preview, “Interaction Models: A Scalable Approach to Human-AI Collaboration,” gives the collaboration idea a more technical shape. Instead of waiting for a complete user turn, the proposed model processes continuous audio, video and text streams. It is designed to handle interruption and overlapping speech, react to visual cues and account for elapsed time. The architecture also pairs a low-latency interaction model with a background model for longer reasoning and tool use, with context shared between them. Thinking Machines describes the approach in its research preview.
Do these 3 things before closing this tab:
1Clear out junk files and repair common Windows errors2Fix the driver behind crashes, sound loss and screen glitches3Repair Windows errors before they cause bigger problemsWhat the Interaction Models preview reports
The preview describes TML-Interaction-Small as a mixture-of-experts model with 276 billion parameters in total and 12 billion active parameters. It uses time-aligned “micro-turns” of roughly 200 milliseconds and supports concurrent input and output. The company reported 0.40 seconds of turn-taking latency on its cited FD-bench measurement and a 77.8 average on FD-bench v1.5. These are company-reported results, not independent rankings or evidence that the system will achieve the same latency under every network, hardware or usage condition.
Thinking Machines described the release as a limited research preview and said a wider release was planned later in 2026. That announcement establishes the plan, not broad availability. The company also identified open challenges including long-session context management, connectivity, alignment, safety and scaling. Continuous voice and video interaction brings its own risks, from accidental activation and background speech to visual privacy, ambiguous social cues and prompt injection through audio or video.
Rank #2
- Built for Local AI Development: AMD Ryzen AI Halo is designed for local AI development and inference, featuring 128GB unified memory and support for up to 200B parameter models to build and run intensive AI workloads locally.
- 128GB Unified Memory: Features 128GB LPDDR5x unified memory at 8000 MT/s with 256 GB/s memory bandwidth, providing a shared memory pool across the CPU, GPU, and NPU to support larger AI models.
- AMD Ryzen AI Max+ 395 Processor: Features 16 cores, 32 threads, and Zen 5 architecture, paired with AMD Radeon 8060S integrated graphics featuring 40 RDNA 3.5 compute units and an AMD XDNA 2 NPU with up to 50 TOPS.
- Linux AI Developer Platform: Purpose-built for Linux-based AI development with full AMD ROCm software support and preloaded tools, models, and workflows optimized for local AI development.
- Compact, Connected Design: Includes a 2TB M.2 SSD, 10GbE LAN, Wi-Fi 7, Bluetooth 5.4, USB-C connectivity, and HDMI 2.1b.
Tinker was the first product: a model-customization platform
Announced on October 1, 2025, Tinker is a platform for researchers and developers who want to fine-tune models without building and operating the full distributed-training infrastructure themselves. Its initial workflows included supervised fine-tuning and reinforcement learning, with early support for Meta’s Llama and Alibaba’s Qwen models. It is not a consumer chatbot: its purpose is to help technical teams customize models and run training workflows. WIRED reported on Tinker’s launch.
The company’s news archive later listed Tinker as generally available and noted vision input on December 12, 2025. WIRED reported that the API was initially free at launch, with charging expected eventually. That was a historical launch signal, not a current price; the available information here does not establish present Tinker pricing.
The underlying bet is that some researchers and organizations want more control than a standard hosted-model API provides: they may want to tune behavior on their own examples, algorithms or domain-specific data without first assembling a training stack. A managed service can lower infrastructure barriers, but it does not remove the need to evaluate model quality, protect sensitive data, monitor outputs or understand the service’s terms. Teams that cannot send sensitive data to a hosted service, or that need a turnkey office assistant rather than training infrastructure, may find Tinker a poor fit.
Inkling: open weights, with limits to what “open” means
In July 2026, Thinking Machines released Inkling, described as its first foundational model. Axios reported that it was built and trained from scratch rather than created by modifying another company’s model, and that full weights were made available through Hugging Face and for fine-tuning through Tinker. The company emphasized customizability rather than claiming Inkling was the strongest general-purpose model. Axios reported on the release.
“Trained from scratch” does not mean the model’s training used no outputs from other models. Axios reported that Inkling’s final training phase used data generated by existing open models, including Moonshot AI’s Kimi K2.5. Those claims can coexist: the model can have its own training run and weights while some training examples are synthetic data produced by other models. The reported facts do not establish benchmark superiority over OpenAI, Anthropic, Google or other frontier labs.
Rank #3
- EVOLUTION RYZEN AI MAX+ 395 MINI PC - GMKtec EVO-X2 is the next evolution in AI mini PC Ryzen Strix Halo series. Thanks to AMD Simultaneous Multithreading (SMT) the core-count is effectively doubled, to 32 threads. Ryzen AI Max+ 395 has 64 MB of L3 cache and can boost up to 5.1 GHz, depending on the workload. The Ryzen AI Max+ 395 is currently rated as the "most powerful x86 APU" on the market for AI computing.
- AI NPU with XDNA 2 ARCHITECTURE - Powered by 16 “Zen 5” CPU cores, 50+ peak AI TOPS XDNA 2 NPU and a truly massive integrated GPU driven by 40 AMD RDNA 3.5 CUs, the Ryzen AI MAX+ 395 is a transformative upgrade and delivers a significant performance boost over the competition. The Ryzen AI Max+ 395 excels in consumer AI workloads like the llama.cpp-powered application: LM Studio. Shaping up to be the must-have app for client LLM workloads, LM Studio allows users to locally run the latest language model without any technical knowledge required and unleash their creativity and productivity.
- AMD RADEON 8090S iGPU GAMING PC - The AMD Radeon RX 8060S offers all 40 CUs with up to 2.9 GHz graphics clock and uses the new RDNA 3.5 architecture. The powerful iGPU is positioned between an RTX 4060 and 4070 laptop GPU and therefore enables gaming in FHD at maximum details in most demanding games. The 8060S can also utilize the full 128GB pool, which is perfect for running LLMs such as Deepseek 70B Q8, which runs comfortably on this machine.
- EIGHT CHANNEL LPDDR5X - LPDDR5X is a new ground breaking memory small form factor installed on-board. With blazing speeds up to to 8000MT/s, it runs 1.5x faster than the DDR5 SODIMMs; 90% better performance over DDR5 SODIMMs in video conferencing and photo editing; 30% better performance in productivity apps; 12% better performance in digital content workloads.
- QUAD SCREEN 8K DISPLAY SUPPORT - EVO-X2 AI Mini PC support 4-screen 4K/8K output via HDMI 2.1 (8K@60Hz), DisplayPort 1.4 (4K@60Hz), and dual USB 4 40Gbps Transfer speed (supporting PD3.0/DP1.4/DATA). Ideal for gaming, video editing, and multitasking, it provides expansive and crisp multi-display support.
Open weights are also not synonymous with a fully open-source or reproducible system. Assess a release by checking what is actually provided: model weights, training code, data, specifications, inference software and license terms. A weights release may allow more control over hosting and customization, but it does not by itself establish access to training data or code, reproducibility, unrestricted commercial use or openness across the company’s other models. Thinking Machines’ launch-era commitment to open scientific work likewise does not mean every model will be open source.
Recommended Free Tools
Funding and NVIDIA infrastructure
In July 2025, Thinking Machines announced a $2 billion seed round at a reported $12 billion valuation. WIRED reported that Andreessen Horowitz led the round, with NVIDIA, Accel, Cisco and AMD among the investors. Contemporary coverage described it as the largest seed round at the time, though such rankings depend on how “seed” is defined and can change. The $12 billion figure is the valuation associated with that financing, not a current market value or an independently audited estimate of what the company is worth. WIRED’s financing report provides the round’s context.
The scale of the financing drew attention because it followed a launch without a public product. It signals investors’ willingness to back the team, ambition and expected compute needs early; it does not establish product-market fit, revenue or customer traction.
On March 10, 2026, Thinking Machines and NVIDIA announced a multi-year strategic partnership. The companies said they intended to deploy at least one gigawatt of next-generation NVIDIA Vera Rubin systems, design training and serving systems for NVIDIA architectures, and support customizable AI at scale. NVIDIA also made a significant investment. The announcement targeted deployment for early 2027; it is a future plan, not evidence that one gigawatt has already been deployed. The company’s announcement sets out the partnership.
That infrastructure could support large-scale training and serving, but it also underscores the capital, hardware availability and energy demands of the company’s strategy. A deep partnership with one supplier can provide scale while increasing dependence on that supplier’s hardware and roadmap.
PC Slower Than It Used to Be?
A free scan shows the junk files, broken settings and background clutter dragging Windows down - then fixes them in one click.Free scan · Windows 10 & 11Outdated Drivers Are Slowing You Down
One free scan finds every outdated or missing driver and matches the right update for your exact hardware.Free scan · exact hardware matchWhat the strategy could mean for users and organizations
Thinking Machines’ visible work points to two related audiences: technical teams that want to customize models and people who may benefit from more continuous multimodal interaction. Plausible applications include domain-specific research models, assistants tuned to a company’s codebase or workflows, scientific and programming tools, and interfaces that can respond to ongoing speech and visual context. Education, design, robotics, operations, meeting assistance and real-time translation are possible areas of use, not confirmed customer deployments.
For a buyer or research team, the practical choice is not simply “open or closed.” It is whether the desired control is worth the added responsibility:
- Hosted closed-model API: often simpler to integrate and operate, but offers less control over model weights and training.
- Managed customization platform: can provide fine-tuning workflows without requiring a team to run a full training stack, but requires review of data handling, service terms, model support and ongoing usage costs.
- Self-hosted open-weight model: offers more control over hosting and adaptation, while requiring suitable GPUs, serving expertise, security and abuse controls, evaluation, license review and ongoing maintenance.
Relevant decision criteria include data sensitivity and residency, fine-tuning methods, latency, multimodal requirements, GPU budget, vendor lock-in, commercial rights, evaluation capacity and enterprise support. The available information does not establish current Tinker pricing, revenue, paying-customer numbers, enterprise contract volume, gross margins or a valuation after the 2025 financing. It also does not establish broad availability for Interaction Models or commercial licensing terms for every Inkling variant.
What remains to be proven
Thinking Machines has progressed from an announced mission to a customization platform, a research preview, a model release and an infrastructure agreement. Those are meaningful milestones, but they answer different questions. Tinker’s availability does not by itself show how many teams use it or whether it is economically sustainable. Inkling’s open weights do not prove it leads general benchmarks. A planned compute deployment does not demonstrate that the hardware is operating, and a research preview does not establish that continuous multimodal interaction is ready for broad production use.
Quick wins for a faster PC:
Repair Windows errors before they cause bigger problemsFix Now →Fix the driver behind crashes, sound loss and screen glitchesFind Drivers →The core business question is whether customization and ongoing interaction create enough value to justify their technical and operational complexity. For customers, that means validating model performance on their own tasks, checking license and data terms, and accounting for deployment and evaluation work—not assuming a company’s mission or financing is proof of a fit.
Quick Recap
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
