Thinking Machines Lab's Inkling is a 975B open-weights model with 41B active parameters, built for developer fine-tuning and full deployment control.
Thinking Machines Lab, the AI startup founded by former OpenAI researchers, released its first model this week. Called Inkling, it's a 975-billion-parameter multimodal system whose full weights are available for anyone to download, modify, and deploy. The model was trained from scratch to process audio, video, and text, as WIRED's Will Knight reported, and its release is designed to help the company carve out space against closed competitors like OpenAI and Anthropic.
What makes Inkling interesting isn't raw benchmark performance. The company itself says Inkling "is not the strongest overall model available today, open or closed" (Thinking Machines Lab). Instead, Thinking Machines is making a deliberate play for developers who want a capable foundation they can reshape for their own purposes, on their own terms. It's a strategic choice that reflects a broader shift in how the AI industry thinks about value creation — and who gets to do it.
What Inkling Actually Is
Inkling is a Mixture-of-Experts (MoE) transformer. That architecture means the model's 975 billion total parameters aren't all active at once. Thinking Machines Lab's announcement notes that only 41 billion parameters are active during any given inference pass, which keeps compute costs and latency lower than a dense model of comparable total size. It supports a context window of up to 1 million tokens and was pretrained on 45 trillion tokens spanning text, images, audio, and video.
The company is also previewing Inkling-Small, a lighter variant with 12 billion active parameters trained using a similar recipe (Thinking Machines Lab). That smaller model targets use cases where cost and speed matter more than peak capability.
Two features stand out. First, Inkling offers what Thinking Machines calls "controllable thinking effort" — the ability to dial reasoning depth up or down depending on the task. The company noted in its blog post that during training, "the chain of thought became more concise over time, dropping grammatical overhead while remaining comprehensible and leaving the final response unaffected" (Thinking Machines Lab). In other words, the model learned to think more efficiently on its own — a phenomenon WIRED highlighted in its coverage.
Second, Inkling is available for fine-tuning on Tinker, the company's customization platform. Thinking Machines built an Inkling Playground into the Tinker console so developers can interact with the model before committing to fine-tuning. The company even demonstrated the concept by having Inkling fine-tune itself — a neat illustration of how AI models are increasingly used to build and improve other AI systems.
Why Open Weights Change the Power Dynamic
To understand why Inkling's release matters, you need to understand what "open weights" actually means in practice. Juncture Policy's glossary on the term defines an open-weight model as giving users "access to the trained artifact, not just the interface." You can download it, run it on your own infrastructure, inspect its behavior, and adapt it for specific needs. That's fundamentally different from calling an API, where you're renting access to someone else's model on someone else's servers under someone else's rules.
This distinction matters for developer autonomy in concrete ways. A startup building a medical imaging tool doesn't want to depend on a cloud provider's pricing whims or content policies. A defense contractor may need to run models in air-gapped environments. A research lab in São Paulo or Nairobi might need to fine-tune a model for languages and contexts that frontier labs in San Francisco never prioritized. Open weights make all of this possible.
Between 2023 and 2026, open-weight models "became a major force in the AI ecosystem as increasingly capable systems were released for local deployment and adaptation," broadening AI competition beyond a handful of hosted frontier services. But that portability also complicates governance, since once weights are public, the original developer has limited control over how the model gets used.
Thinking Machines is leaning into this tradeoff rather than shying away from it. The company's stated mission is "to build AI that extends human will and judgment," and releasing full weights is the most literal expression of that philosophy — handing the artifact over and trusting developers to shape it (Thinking Machines Lab).
The Competitive Landscape for Open Models
Inkling doesn't exist in a vacuum. Meta's Llama family has been the most prominent open-weights effort from a major tech company. And the investment flowing into this space is substantial. WIRED reports that Nvidia plans to spend $26 billion over five years building open-weight AI models, a move that could see the chipmaker evolve from infrastructure provider to frontier lab competitor. Nvidia's strategy is particularly interesting because its open models are tuned to its own hardware, creating a flywheel that reinforces GPU sales.
DeepSeek, the Chinese AI lab, has also demonstrated that open-weight models can achieve frontier-level performance (DeepSeek Transparency Center). The competitive pressure from these players has made it harder for closed-model companies to justify keeping weights locked down purely on capability grounds.
Inkling enters this landscape as a smaller player with a different pitch. Where Meta and Nvidia bring scale and distribution, Thinking Machines is emphasizing the combination of multimodal capability, efficient reasoning, and a purpose-built fine-tuning platform. It's a bet that developer experience and customization tooling matter as much as raw model size.
The infrastructure to support these deployments continues to expand. As we covered in our reporting on Meta's $10 billion Tulsa data center, the buildout of AI-optimized compute facilities is accelerating across the country. That growing infrastructure base makes it more practical for organizations to run large open-weight models locally or in private cloud environments, rather than depending on a single provider's API.
What Inkling Means for Developers Building AI Applications
The practical implications of Inkling's release play out at several levels.
Customization without permission. Developers can fine-tune Inkling for domain-specific tasks without negotiating enterprise contracts or waiting for a provider to add features. The Tinker platform lowers the barrier further by handling the infrastructure complexity of fine-tuning a model with hundreds of billions of parameters.
Cost predictability. Running your own model means your inference costs are tied to your compute budget, not a per-token pricing schedule that can change quarterly. For applications processing millions of requests, this can be the difference between a viable business and a money pit.
Multimodal flexibility. Inkling's native support for text, images, audio, and video in a single model simplifies architectures that would otherwise require stitching together multiple specialized systems. For developers building applications that need to reason across media types — think accessibility tools, content moderation, or industrial inspection — this is a meaningful reduction in complexity.
The governance question. Open weights also mean developers bear more responsibility. Without a provider's safety filters as a backstop, teams deploying Inkling need their own guardrails. That's a feature for sophisticated organizations and a risk for less careful ones.
Where This Goes Next
Thinking Machines has positioned Inkling as "just the start" of a model family it plans to continue developing. The Inkling-Small preview suggests the company is already thinking about the spectrum of deployment scenarios, from resource-constrained edge devices to full-scale cluster deployments.
The broader trajectory is clear. Open-weight models are no longer a niche concern for researchers. They're becoming a mainstream option for production AI applications, backed by billions in investment from companies like Nvidia and Meta. Thinking Machines is betting it can compete not by building the biggest model, but by building the most useful one to customize.
For developers, the calculus is straightforward. More capable open-weight models mean more choices, less lock-in, and greater control over the AI systems they build. The tradeoff is taking on more responsibility for safety, alignment, and governance. That's a deal many teams will gladly take.
