The 975-billion-parameter system prioritises customisation and efficiency as the startup challenges the closed platforms dominating advanced artificial intelligence
Mira Murati’s Thinking Machines Lab has released its first general-purpose artificial-intelligence model, marking the heavily financed startup’s most significant attempt yet to challenge the dominance of OpenAI, Anthropic and other leading developers.
The model, called Inkling, is being distributed under an “open weights” approach. This allows organisations and developers to download the model’s trained parameters, operate it on their own infrastructure and adapt it using proprietary information rather than relying exclusively on a vendor-controlled online service.
Thinking Machines is positioning Inkling less as the most powerful model available and more as a flexible foundation that companies can reshape for specialised tasks. The startup has acknowledged that the system does not outperform every leading closed model, instead emphasising its balance across different subjects, adaptability and lower operating costs.
Inkling contains 975 billion parameters, although only about 41 billion are activated for a given task. That architecture is designed to reduce the computing power needed to run the model while preserving the broader knowledge stored across the full system.
Its release represents a strategic bet that the next phase of the AI industry will not be determined solely by which company produces the highest-scoring general model. Thinking Machines believes developers and businesses will increasingly value systems they can modify, control and deploy according to their own requirements.
That approach contrasts with the tightly controlled platforms offered by many of the industry’s largest laboratories. Companies using proprietary systems generally send requests through application-programming interfaces and receive outputs without gaining access to the underlying model weights.
Closed models can offer strong performance, regular upgrades and managed security, but they also leave customers dependent on the provider’s prices, policies and technical infrastructure. Open-weight alternatives give users greater control, particularly when handling sensitive corporate information or building highly specialised applications.
Thinking Machines was founded in 2025 by Murati, the former chief technology officer of OpenAI, alongside a team that includes researchers and engineers with experience at several major AI companies. The San Francisco-based startup raised approximately $2 billion in seed financing at a reported valuation of $12 billion before releasing a general-purpose model.
The size of that investment created enormous expectations. Inkling will now serve as an early test of whether Murati’s company can convert its high-profile leadership, research talent and substantial computing resources into technology that attracts a broad developer and enterprise audience.
The model builds on Thinking Machines’ earlier product, Tinker, a platform that helps researchers and companies fine-tune open models without managing the full complexity of distributed computing infrastructure. The company says Tinker gives users control over their data and training methods while handling much of the technical work required to customise large systems.
Inkling will be available through Tinker and other developer platforms, making the model part of a larger ecosystem focused on adaptation rather than a single consumer chatbot.
The strategy has already attracted interest from large institutions. Bridgewater Associates used Tinker to customise Alibaba’s Qwen model for financial work, reportedly improving performance on certain tasks while reducing processing expenses. The collaboration offers an example of the market Thinking Machines hopes to capture: organisations that do not necessarily want to build a foundation model from the beginning but require more control than standard commercial chatbots provide.
Inkling also arrives during a period of growing competition between American and Chinese developers in the open-model sector. Chinese laboratories have gained significant support among developers by releasing capable, customisable systems, while some Western companies have concentrated on proprietary products.
Reuters reported that many businesses increasingly turned to Chinese models after Western open-model efforts lost momentum, particularly following disappointment surrounding Meta’s Llama 4. Thinking Machines is attempting to provide a major American alternative within that market.
The company’s model nevertheless faces significant obstacles. OpenAI, Anthropic and Google continue to lead many overall performance benchmarks, supported by enormous computing budgets, established distribution networks and rapidly expanding product ecosystems.
Inkling’s size could also limit the number of organisations capable of running it independently. Even with only a fraction of its parameters active at once, deploying a model of this scale requires substantial technical expertise and expensive hardware.
Thinking Machines has sought to address those concerns through a partnership with Nvidia, whose processors were used in developing Inkling. The two companies have also announced a broader multiyear computing agreement intended to provide the startup with access to future generations of AI infrastructure.
Open weights introduce additional safety questions. Once a model has been downloaded, its original developer has less ability to monitor how it is modified or prevent safeguards from being removed. Thinking Machines says Inkling underwent testing intended to reduce the risk of misuse, although the company has acknowledged that greater openness inevitably transfers more responsibility to users.
Supporters argue that open models encourage competition, reduce dependence on a small group of corporations and allow universities, governments and businesses to inspect and adapt important technology. Critics warn that the same accessibility can make powerful systems easier to exploit for cyberattacks, disinformation or other harmful activities.
Murati’s company has framed the debate in broader political and economic terms. In a manifesto accompanying the launch, Thinking Machines criticised the concentration of AI development within a handful of institutions and argued that useful systems should incorporate the specialised knowledge held by individual organisations and communities.
The release therefore represents more than the introduction of another large model. It is an attempt to promote a different structure for the AI industry—one in which businesses own and customise more of the technology they use rather than accessing intelligence solely through closed corporate platforms.
Inkling is unlikely to displace the leading commercial models immediately. Its importance will instead depend on whether developers conclude that greater control and adaptability can compensate for any gap in general performance.
For Thinking Machines, that is the central wager: the future of artificial intelligence may belong not only to the company with the smartest model, but also to the one that gives others the greatest freedom to reshape it.
