Thinking Machines Lab, the AI startup founded by former OpenAI chief technology officer Mira Murati, released its first in-house model on July 15, 2026. Called Inkling, it is — by the company's own candid admission — open-weight, enormous, and not the best model on the market. That framing is deliberate, and it reveals a distinct strategy in a field obsessed with leaderboard supremacy.
What Inkling Is
Inkling is a Mixture-of-Experts transformer with 975 billion total parameters and 41 billion active per token. 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 model reasons natively across text, images, and audio, and balances cost against performance through what the company calls efficient, controllable "thinking effort." For now, its outputs are limited to text — though that includes code, styled artifacts, and structured data.
Crucially, it ships under a permissive Apache 2.0 license. The full weights are available on Hugging Face, and the model is live for fine-tuning on Tinker, the company's customization platform. Inkling is also the first of a family: alongside it, Thinking Machines previewed Inkling-Small, a 276-billion-parameter model with 12 billion active, whose weights will follow once testing completes.
A Deliberately Different Bet
Thinking Machines is refreshingly blunt that Inkling is not the strongest model available, open or closed. Instead, the company is betting that enterprises ultimately care less about the smartest general-purpose model than about one they can make their own. The pitch rests on three pillars: multimodal understanding, efficient reasoning, and easy fine-tuning through Tinker, where customers own whatever they build — and carry the safety risk of it.
Efficiency is a headline claim. The company says Inkling uses roughly a third as many tokens as Nvidia's latest open-weight Nemotron 3 Ultra to reach the same coding performance. In a world where inference cost increasingly determines whether an AI deployment is economical, token efficiency is a competitive weapon, not a footnote.
The lab also emphasizes a design focus it calls epistemics — training the model for calibration, instruction following, and resistance to censorship. A useful model, the team argues, should express the right amount of confidence in its answers, including on questions that are not yet settled. That is a pointed contrast to models tuned to sound authoritative regardless of certainty.
Not a Revenue Play
Notably, Thinking Machines is not trying to monetize Inkling directly. The company's revenue comes from Tinker, the fine-tuning tool it sells to customers such as hedge fund Bridgewater Associates to sharpen AI on specialized tasks. The model, in effect, is a high-quality base designed to funnel customization work onto its paid platform.
On engineering pedigree, the model was built from scratch on Nvidia's latest AI infrastructure rather than adapted from another lab's weights — underscoring a deep partnership with the chip giant. Thinking Machines did acknowledge using other open models, including Moonshot AI's Kimi K2.5, to help generate some early post-training data before large-scale reinforcement learning took over.
Why It Matters
Inkling matters less for where it ranks than for what it represents. The independent analyst Simon Willison observed that it gives the US open-weights ecosystem a genuine new contender to sit alongside Nvidia's Nemotron and Google's Gemma — an ecosystem that has increasingly leaned on strong open models from Chinese labs. For organizations wary of dependence on either closed APIs or foreign open weights, a permissively licensed, US-built, customizable base is strategically valuable.
The release also reframes what "winning" in AI can mean. The dominant narrative pits labs in a sprint for the top of every benchmark. Thinking Machines is arguing that the real enterprise need is a model you can adapt, audit, and own — and that a slightly-less-capable-but-highly-customizable model may beat a frontier system that cannot be tailored to a specific domain.
Finally, there is the sheer speed of execution. TechCrunch noted that OpenAI took about five years to ship and earn, and Anthropic roughly three; Murati's lab says it went from founding to a from-scratch frontier-class open model in about nine months. That pace, backed by top-tier talent and Nvidia infrastructure, signals just how compressed the timeline for building a serious AI lab has become — and how quickly a well-funded newcomer can put real weights on the table.
Whether enterprises embrace the customization-first thesis will take months to judge. But Inkling is a clear, credible marker that the open-weights race now has another serious American runner.
