Some links on this page are affiliate links: if you buy through them we may earn a commission, at no extra cost to you.
Former OpenAI chief technology officer Mira Murati launched Thinking Machines Lab on February 18, 2025, with roughly 30 researchers and engineers recruited from OpenAI, Meta, Mistral and other AI companies. The company is no longer just a high-profile stealth startup: it now offers Tinker, a fine-tuning platform, and the open-weights multimodal models Inkling and Inkling-Small.
Its clearest strategy is to let researchers and organizations customize advanced AI, rather than simply sell another general-purpose chatbot. That makes Thinking Machines a potential frontier-model competitor to OpenAI, but a specialized one whose commercial success will depend on developer adoption, infrastructure economics and reliable production support.
What Mira Murati launched
Thinking Machines Lab is a combination of AI research lab, frontier-model developer and model-customization company. Murati is its CEO and co-founder. At launch, the company said it wanted to build AI systems that adapt to human expertise, work more naturally with people and make advanced AI more understandable and useful. The company later described its mission as building AI that extends human will and judgment through customization, human participation and distributed alignment.
Recommended Free Tools
The launch announcement did not name a model, publish a product, disclose a technical architecture or provide confirmed financing terms. That secrecy described the company’s public stage in February 2025; it does not describe its position by August 2026.
#1 Best Overall
TechCrunch’s launch report and the company’s mission statement provide the original framing.
The team behind the company
Thinking Machines’ early credibility came from concentrating experienced model researchers and engineers in one new organization. The launch team was reported at about 30 people, including recruits from OpenAI, Meta, Mistral, Character.AI and Google DeepMind.
| Person | Thinking Machines role | Previously associated with |
|---|---|---|
| John Schulman | Chief scientist; OpenAI co-founder | OpenAI |
| Barret Zoph | Chief technology officer | OpenAI vice president of research |
| Andrew Tulloch | Chief architect | OpenAI researcher |
| Lilian Weng | Research and technical leadership | OpenAI safety and robotics |
| Luke Metz | Research and post-training | OpenAI |
| Alec Radford | Adviser | OpenAI researcher |
| Bob McGrew | Adviser | OpenAI chief research officer |
The company’s public team page distinguishes its leadership, employees and advisers. Not everyone associated with the launch was an OpenAI executive; many were researchers, engineers or technical staff. The recruiting was significant because frontier AI depends on a small pool of people who have trained and deployed large models. It also creates risks: a celebrated roster does not guarantee cohesion, retention or a product that customers will pay for.
From stealth startup to Tinker
Thinking Machines introduced Tinker on October 1, 2025. It is a managed API for fine-tuning language models: customers control their data and training logic while the service handles distributed training, scheduling, resource allocation and failure recovery.
Rather than exposing only a high-level “send a prompt” interface, Tinker provides lower-level operations such as forward_backward, optim_step, sample and save_state. Its LoRA-based approach shares compute while allowing a team to train an adaptation of a base model. Tinker began as a private beta, reached general availability in December 2025, and added OpenAI-compatible sampling and vision-input support.
This is aimed at researchers, startups, universities and AI teams that need control over post-training but do not want to build a distributed-training system. It is not a consumer chatbot and is a poor fit for a buyer seeking a turnkey enterprise application.
Inkling and Inkling-Small put customization at the center
On July 15, 2026, Thinking Machines released Inkling, its first open-weights model. The specifications below are company disclosures, not independent audits.
The Tool Desk
Outbyte Driver Updater FREEFix the driver behind crashes, sound loss and screen glitchesFind Drivers →Outbyte PC Repair FREEClear out junk files and repair common Windows errorsFree Scan →| Model | Total parameters | Active parameters | Context | Modalities |
|---|---|---|---|---|
| Inkling | 975 billion | 41 billion | Up to 1 million tokens; Tinker options list 64K and 256K | Text, image and audio input |
| Inkling-Small | 276 billion | 12 billion | Up to 1 million tokens | Native audio and image reasoning |
Both use mixture-of-experts designs. Thinking Machines says Inkling was trained on 45 trillion tokens spanning text, images, audio and video, and can be fine-tuned through Tinker. The company explicitly does not present Inkling as the strongest overall model. Its proposed advantages are open weights, multimodality, controllable reasoning effort and customization.
Inkling-Small, released July 30, 2026, is intended to deliver comparable capabilities at substantially lower compute cost. Open weights improve portability and control, but they also transfer deployment, security, monitoring and inference responsibilities to the buyer.
Funding, valuation and planned compute
TechCrunch reported a $2 billion seed financing in 2025. Separate reports put Thinking Machines’ valuation at $10 billion and $12 billion, so the defensible description is a reported valuation range rather than one confirmed figure. That would make it one of Silicon Valley’s largest reported seed financings, but not a verified record without a common ranking method.
In March 2026, Axios reported a long-term NVIDIA partnership involving gigawatt-scale compute beginning in 2027. The announcement signals frontier-scale ambitions, but announced future capacity is not the same as delivered or currently available compute.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
Capital and infrastructure can fund large-model training, hiring and product development. They do not prove model superiority, customer demand or sustainable margins.
Is Thinking Machines a direct OpenAI competitor?
Yes, in the broad strategic sense: it has senior OpenAI alumni, substantial reported financing, its own models and plans for large-scale compute. But calling it an OpenAI replacement oversimplifies the company’s current position.
| Dimension | Thinking Machines’ position | What that means |
|---|---|---|
| Model ownership | Develops its own models | It is more than an application wrapper. |
| Openness | Inkling models use open weights | Customers can seek more control and portability, subject to licensing. |
| Customization | Tinker exposes fine-tuning workflows | Researcher control is the central differentiator. |
| Consumer access | No general consumer chatbot is described here | The clearest offering is developer and research infrastructure. |
| Production maturity | Serverless inference remains in beta | Buyers should not assume intensive production workloads are supported. |
The company’s model is closer to “frontier models plus adaptation tools” than to a single ChatGPT-style product. It competes with closed-model providers, cloud platforms and open-model ecosystems, while targeting users who value control over convenience.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.What users and businesses can actually buy
Tinker
Sign-up information is available at thinkingmachines.ai/tinker, with documentation and model pricing at the Tinker model documentation. Pricing is usage-based per million tokens, and checkpoint storage is listed at $0.10 per GB-month. The documentation reviewed for August 2026 displayed limited-time discounts, including $5.61 per million tokens for Inkling 64K training and $1.73 for Inkling-Small 64K training. These are volatile listed prices, not permanent rates.
Windows Errors? Fix Them Before They Spread
Repair common Windows errors and clear accumulated junk for a smoother, more stable PC - no reinstall needed.Free scan · no reinstallCrashes, No Sound, or Screen Glitches?
Random freezes, missing sound and display glitches usually trace back to one bad driver. Find and replace yours safely.Free scan · under a minuteTinker’s documentation says serverless inference is in beta and is not recommended for intensive production use until it leaves beta. That makes it more suitable for experimentation, research and controlled workloads than for an organization demanding mature service-level guarantees.
Inkling models
The Inkling model page and company announcements describe open-weights models that can be fine-tuned, self-hosted or accessed through infrastructure providers such as Together AI, Fireworks, Modal, Databricks, Baseten and Hugging Face. Availability, licensing, hosted pricing and support vary by provider.
Inkling is a plausible fit for organizations with machine-learning infrastructure, evaluation expertise and a clear need for domain adaptation. A small business without GPU capacity or model-operations staff may spend less time and money using a hosted general-purpose model.
Questions a prospective buyer should answer
- Is the required model available in the organization’s region?
- Can training data be retained, deleted or isolated as privacy rules require?
- What service-level commitments apply to training and inference?
- Can fine-tuned checkpoints be exported or moved to another provider?
- What will inference cost after training, at the expected traffic level?
- Does the license permit commercial redistribution, embedding or self-hosting?
- Is the team prepared to evaluate long-context, multimodal and safety performance on private data?
- Is the workload appropriate for a beta service, or does it require production-grade support now?
The business risks behind the headline
- Capital intensity: Training and serving very large models require continuing compute and infrastructure spending.
- Product-market fit: Researchers may want low-level controls while mainstream enterprises prefer simpler managed APIs.
- Open-weight economics: Wider adoption can come with less direct model-licensing revenue.
- Competition: OpenAI, Anthropic, Google, Meta, Alibaba, DeepSeek and other providers are improving models and customization tools.
- Evidence quality: Parameter counts, training-data volumes and benchmark claims from the company should be treated as company-reported until independently reproduced.
- Technical limits: Fine-tuning is not automatically cheaper than prompting, retrieval augmentation or distillation; a million-token context window does not guarantee reliable long-context reasoning.
What the talent story means now
Murati’s recruiting haul explains why Thinking Machines attracted immediate attention: it assembled people with scarce experience building frontier systems before showing a public product. Tinker and the Inkling releases now provide a more concrete explanation of the company’s strategy. It is trying to turn that talent and capital into a platform where advanced models can be adapted to specific users and organizations.
Whether customization becomes a large, durable business remains unresolved. The company has the financing, technical credentials and announced compute ambitions to compete at the frontier, but those advantages must translate into dependable products, repeat usage and sustainable operating economics.
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.
