In a move that has rattled the closed-AI establishment, Mira Murati’s startup Thinking Machines Lab released its first AI model this week — a 975-billion-parameter system called Inkling — and made the weights publicly available for anyone to download, modify, and deploy. For the person who spent years as CTO of OpenAI, the company synonymous with keeping its best models behind a paywall, the decision is more than a product launch. It’s a statement of intent.
What Took Nine Months to Build
Thinking Machines was founded in early 2025 by Murati alongside fellow OpenAI veterans John Schulman and Lilian Weng. The company has operated largely out of public view for 18 months, quietly assembling infrastructure and a team of roughly 200. Inkling is its first proof point.
The model is a Mixture-of-Experts (MoE) architecture with 975 billion total parameters, but it only activates about 41 billion of those for any given task — a design that keeps inference fast and cost-effective without shrinking the model’s overall capacity. It was trained on 45 trillion tokens spanning text, images, audio, and video, and is built to reason natively across all four modalities.
What Thinking Machines is proudest of may not be the architecture itself but the speed. OpenAI needed roughly five years to bring a frontier-class product to market. Anthropic took about three. Thinking Machines says it did it in nine months — a timeline that, if accurate, resets expectations for how quickly a well-funded, experienced team can stand up a competitive lab from scratch.
The company trained Inkling entirely on Nvidia GB300 NVL72 systems, part of a strategic partnership with Nvidia struck in March to deploy a gigawatt of Vera Rubin computing capacity. That partnership is also a signal: even a startup with elite talent and a novel approach still needs to be close to the hardware supply chain to compete.
A Different Kind of Model
Thinking Machines is deliberate about what Inkling is not. The company’s own materials state that it is “not the strongest model available today, closed or open.” It doesn’t aim to top every benchmark. What it aims to do is be reliably calibrated — flagging uncertainty rather than confabulating, letting users dial “thinking effort” up or down to trade reasoning depth for speed, and costing meaningfully less to run than comparable closed alternatives.
One internally reported benchmark is striking: Inkling uses a third as many tokens as Nvidia’s Nemotron 3 Ultra to reach the same coding performance. If that holds up under independent evaluation, it means running Inkling costs roughly a third as much for coding tasks — a number that matters enormously at enterprise scale.
The model is positioned primarily as a starting point rather than a finished product. Thinking Machines paired the release with Tinker, its model-customization platform designed to let organizations fine-tune Inkling on their own data. The core argument is that a model shaped by a company’s proprietary knowledge will outperform a general-purpose model bought by subscription — and that the lab selling subscriptions has an incentive to absorb the training signal you generate, not return it to you.
That argument has been gaining traction beyond Thinking Machines’ own walls. In a blog post published earlier this month, Microsoft CEO Satya Nadella warned that enterprises using proprietary AI models effectively pay twice: once in subscription fees, and again by handing over business knowledge that can be incorporated into future model versions. Hugging Face CEO Clem Delangue has predicted that frontier models will increasingly be reserved for experimentation while production AI work shifts to private or open-source alternatives. Inkling is a bet that those predictions are correct.
The Open-Weight Moment
Thinking Machines didn’t release Inkling into a vacuum. The past week has been, arguably, the biggest stretch in the young history of open-weight frontier AI.
China’s Moonshot AI released Kimi K3 — a 2.8-trillion-parameter model that topped the coding arena at Arena.ai with a 76 percent pairwise win rate, beating Claude Fable 5 — with open weights promised by July 27. DeepSeek V4, which has led open-weight leaderboards for months, is due for a stable release on July 24. And now Inkling, from a team with impeccable American frontier credentials, has joined them.
Four of the most significant model releases of July 2026 are open-weight. That’s not a coincidence. It reflects a broad shift in how the most capable AI actors think about where competitive advantage actually lives. If models are becoming commodities — something increasingly difficult to dispute as new frontier-class releases arrive every few weeks — then the moat isn’t the model itself. It’s the infrastructure around it, the tools for customization, the data that shapes it, and the enterprise relationships that generate both.
Thinking Machines is betting that Tinker, not Inkling, is where its revenue comes from. The model is free. The customization platform, the training runs, the hosting ecosystem built around it — those are the business.
Evidence It Can Work
The most compelling data point Thinking Machines has offered is a joint project with Bridgewater Associates, the world’s largest hedge fund. Researchers from both companies took an open-source base model and fine-tuned it on Bridgewater’s own financial expertise. The resulting model scored 84.7 percent on financial reasoning benchmarks — beating top proprietary models — at roughly one-fourteenth the operational cost.
The caveat is real: those results come from the two companies’ own evaluation, not an independent one. But even directionally, a domain-specialized model outperforming general-purpose frontier systems while costing a fraction as much is exactly the thesis Thinking Machines is selling. If the numbers survive independent scrutiny, the case for open-weight customization over closed-model subscriptions becomes substantially harder to dismiss.
The Distillation Question
Not every aspect of Inkling’s origins is clean. Thinking Machines acknowledges it used outputs from other open-weight models — including Moonshot AI’s Kimi K2.5 — to generate some of its early post-training data before large-scale reinforcement learning took over. The company says the next model will use fully self-contained post-training.
The practice of training on outputs from other models, known as distillation, has drawn legal scrutiny across the industry as courts begin to examine what’s permissible when a model learns from another model’s outputs rather than from original human-generated data. Thinking Machines’ use of open-weight models for this purpose keeps it on firmer legal ground than distilling from proprietary systems, but the question of where the lines are is not yet settled.
What It Means for the Industry
The open-weight offensive of July 2026 is a forcing function. If developers can download Inkling, Kimi K3, and DeepSeek V4 for free — and get within striking distance of GPT-5.6 and Claude Fable 5 on the workloads they care about — the closed labs face a harder question than they’ve had to answer before. Not “is our model better?” but “is our model enough better to justify the cost?”
For enterprises with serious ML talent, the answer is increasingly no. For enterprises without that talent, closed models with integrated tooling still hold real advantages. But the talent question is a race condition: as open-weight ecosystems mature, the expertise required to run and customize them will become more available, not less.
Mira Murati left OpenAI in September 2024. Less than two years later, the company she built has a frontier-class model and is giving it away. Whatever else Inkling turns out to be, it’s a credible opening argument in a debate that will define the second half of the decade: who controls AI, who profits from it, and whether the best models need to be owned at all.
Sources: TechCrunch · The Decoder · SiliconAngle · The Register · VentureBeat
