Thinking Machines Enters the A.I. Race With Inkling

Thinking Machines Unveils Inkling, Its First Model, in a Bid to Carve Out a Place in the A.I. Race

Thinking Machines Lab, the artificial intelligence startup founded by former OpenAI leaders including Mira Murati, on Tuesday released its first foundation model, a sprawling open-weights system called Inkling that can process text, images and audio and is aimed less at winning headline benchmarks than at giving companies a model they can adapt for their own use.

The company said Inkling is a 975-billion-parameter mixture-of-experts model with 41 billion active parameters at a time, a design intended to deliver some of the scale of the largest systems without requiring every part of the network to run on every query. It is being released under the Apache 2.0 license, with a context window of up to one million tokens, placing it among a growing class of models built to ingest unusually large amounts of material in a single session.

The launch is a notable moment for Thinking Machines, which was founded in 2025 and has spent much of the past year attracting talent, building computing capacity and introducing Tinker, a platform for customizing models. Until now, the company had not shipped a flagship model of its own. Inkling gives the startup a concrete product in a field where credibility is increasingly tied not just to pedigree and fundraising, but to what a lab can actually train, release and support.

In announcing the model, the company was unusually direct about its ambition. Inkling, it said, is not the strongest model available, whether open or closed. Instead, it is being pitched as a base model that enterprises can fine-tune, control and deploy with more freedom than they would get from proprietary systems offered by rivals like OpenAI, Anthropic and Google.

That positioning reflects a broader shift in the A.I. market. For much of the past several years, the industry revolved around a simple contest over who had the most capable general model. But as systems have improved and costs have mounted, many companies have started to prioritize other features: how cheaply a model can run, how easily it can be tuned for a specialized task, whether it can be hosted privately, and how much latency and “reasoning” effort a user can dial up or down.

Thinking Machines is leaning heavily into that last idea. The company says Inkling offers controllable “thinking effort,” allowing developers to trade off speed and cost against deeper reasoning. For enterprise customers, that promise could be practical rather than philosophical: a customer-service workflow may need quick, cheap responses, while legal review or software debugging may justify slower, more expensive inference.

An Open-Weights Bet

Inkling’s release also lands in a strategically important gap in the market. While open models have proliferated, the strongest and most visible alternatives have often come either from Chinese labs or from American companies that keep their leading systems closed. Thinking Machines is betting there is room for a large Western open-weights model that companies can inspect, modify and integrate into their own infrastructure.

That may matter especially to businesses wary of handing sensitive data or core operations to an external model provider. An open-weights release does not make a model easy to run — a system of Inkling’s size will still demand substantial hardware, even with mixture-of-experts efficiencies and future quantized variants — but it does give customers more control over where and how it is deployed.

The company has tried to smooth that path by launching Inkling with support from an array of infrastructure and tooling partners, including providers of inference, fine-tuning and serving frameworks. It also made the model available for immediate experimentation through its Tinker platform and a new playground for testing.

Still, openness alone is unlikely to decide the contest. The central question is whether Inkling is good enough, cheap enough and flexible enough to pull developers away from stronger closed systems or from leading open models already circulating in the market.

A Crowded July for A.I. Labs

The debut comes during an unusually compressed stretch of major model announcements. In recent weeks, the industry has seen a steady churn of new releases, leaks and upgrades from some of the world’s largest labs, including OpenAI, Meta, xAI and ByteDance, with more expected in the weeks ahead. Rather than one clear leader, the market is increasingly defined by specialization: longer context windows, stronger coding performance, better multimodal input, more agentic behavior and sharper pricing tiers.

Inkling fits squarely into that moment. Its headline features — multimodality, one-million-token context, open weights and customization — are less a declaration of outright dominance than an attempt to meet what developers now say they need.

The company has drawn particular attention to audio, an area that has often lagged behind text and image capabilities in mainstream models. It says Inkling was trained to understand video and audio and has shown strength on audio-focused benchmarks, a sign that multimodal systems are increasingly being judged not just on whether they can accept many forms of input, but on whether they can reason across them competently.

Why It Matters Now

For Thinking Machines, the release is as much about proof as performance. In March, the company announced a multiyear partnership with Nvidia to deploy at least one gigawatt of next-generation Vera Rubin systems, signaling ambitions on a frontier scale. Inkling is the first visible output of those ambitions: evidence that the startup can move from recruiting and infrastructure to a product developers can use.

For the broader industry, the model underscores a changing balance of power. The next phase of the A.I. race may not be won solely by the lab with the highest benchmark score. It may hinge on who offers the most useful mix of performance, deployability and control.

That does not mean Inkling is guaranteed traction. Large mixture-of-experts models remain expensive and operationally complex. Enterprises may still prefer the convenience and perceived quality of closed offerings. And the open-model field remains fiercely competitive, with Chinese developers in particular moving quickly.

But the release gives Thinking Machines a foothold where, until now, it had mostly promise. In a market increasingly crowded with giant models, that alone is significant: a startup that had been known primarily for its founders and ambitions now has a model of its own — and a clear argument for why some customers might want it.

Sources

Further reading and reporting used to add context:

AI News

Thinking Machines Enters the A.I. Race With Inkling