Thinking Machines Makes an Unusual Admission
Thinking Machines has done something almost no AI lab ever dares. When most companies launch a model they insist it is the best, but on July 15 the secretive startup founded by former OpenAI chief technology officer Mira Murati released Inkling, its first model trained entirely from scratch, and told the world plainly that it "is not the strongest overall model available today, open or closed." That candor is not modesty. It is the whole strategy, and it explains why Thinking Machines is trying to win a game everyone else is playing backwards.
The company has spent roughly eighteen months in near silence since Murati walked out of OpenAI, raising about $2 billion at a $12 billion valuation and hiring a roster of researchers from the biggest labs. Expectations were enormous. What arrived is not a bid to top a leaderboard. It is a 975-billion-parameter foundation model designed to be pulled apart, retrained, and rebuilt by whoever downloads it. In a market obsessed with a single headline benchmark number, that is a genuinely contrarian wager.
Inside Inkling
Inkling is a Mixture-of-Experts transformer with 975 billion total parameters, of which only about 41 billion activate on any single token. It supports a context window of up to one million tokens and was pretrained on 45 trillion tokens spanning text, images, audio and video, reasoning natively across all four. Alongside it, the lab previewed Inkling-Small, a lighter model that fires just 12 billion parameters yet matches or beats its larger sibling on a surprising number of tests.
The full weights are public under a permissive Apache 2.0 license, hosted on Hugging Face, which makes Inkling one of the largest fully open American models ever shipped. On the benchmarks the company chose to publish, it lands respectably rather than spectacularly. It scores 97.1 percent on the AIME 2026 math exam, 87.2 percent on GPQA Diamond, and 77.6 percent on SWE-bench Verified. Frontier closed systems from Anthropic and OpenAI still post higher numbers on the hardest reasoning tasks, and the company does not pretend otherwise.
Inkling's distinguishing trick is efficiency rather than raw peak performance. Developers can dial its "thinking effort" up or down to trade accuracy against speed and cost, and the lab says the model reaches the same score as Nvidia's Nemotron 3 Ultra on agentic coding while spending roughly a third of the tokens. For a model meant to run millions of times inside long workflows, that cost curve matters more than a leaderboard trophy. The system also reasons over audio and vision without a separate encoder, a design choice pointing toward the real-time, voice-and-vision assistant Thinking Machines has hinted it wants to build.
The real product is Tinker
To understand why a lab would ship a model it openly ranks below the competition, look at how it makes money. Thinking Machines does not sell Inkling as a finished chatbot. It sells Tinker, a platform for fine-tuning open models on a customer's own data, and Inkling is now the flagship base that platform is built around. The bet is that AI a company can bend to its own knowledge will outperform the one-size-fits-all systems the giants rent out by the token.
There is early evidence this works. Researchers at Thinking Machines and the hedge fund Bridgewater Associates took an open model and trained it further on Bridgewater's financial expertise. The customized result reportedly scored about 84.7 percent on financial reasoning tests, beating top proprietary models while costing roughly a fourteenth as much to run. That is the pitch in one number: a "second-best" general model, tailored, can beat a first-place general model on the task you actually care about, for a fraction of the price.
To drive the point home, Thinking Machines had Inkling fine-tune itself. Using Tinker, the model wrote its own training job, ran it, evaluated the result, and loaded the new weights back into its coding harness, teaching itself a quirky target behavior that prompting alone could not reliably produce. It is a marketing demo, but also a clear statement of intent. The value is not the model. It is the ability to change the model.
An open-weights war heats up
Inkling did not land in a quiet room. The same week, Chinese lab Moonshot released Kimi K3, an even larger open model that rattled global markets and dragged AI and chip stocks lower as investors relived last year's "DeepSeek moment." Mira Murati's former colleague Ilya Sutskever, along with Meta, Alibaba, Nvidia and a growing field of Chinese labs, has helped turn open weights from a fringe cause into the industry's most contested battleground. The question is no longer whether capable models will be freely downloadable. It is whose values, safety choices and business model will ride along with them.
On safety, Thinking Machines is trying to make openness defensible. It reports that Inkling posts the strongest built-in safeguards of any open model it compared on FORTRESS, a benchmark testing refusal of weapons and violence requests, while avoiding over-refusal of harmless look-alike questions. The lab commissioned outside testers for dangerous-capability evaluations covering cyber, biological and loss-of-control risks. Whether those guardrails survive contact with fine-tuning, the very thing the company is selling, is an open research problem it admits it is still studying.
There is also a geopolitical subtext. A large, permissively licensed, safety-tested American open model gives Western developers an alternative to leaning on Chinese releases like Kimi and DeepSeek for their open-weights foundation. In that sense Inkling is as much a piece of soft infrastructure as a product, and its ecosystem partners, from Hugging Face to Together AI and Databricks, are lining up to distribute it.
Why a "second-best" model might win
The temptation is to score Inkling against GPT-5.6 and Claude Fable 5 and conclude that Thinking Machines fell short of the hype. That reading misses the point. Murati is not selling the smartest oracle. She is selling raw material and the tools to shape it, wagering that the next phase of the industry belongs to organizations that own and adapt their models rather than rent access to someone else's. If she is right, the relevant benchmark is not who tops a math exam this quarter but who gives developers the most useful clay.
It is a strategy with real risks. Open weights are expensive to produce and hard to monetize directly, the fine-tuning market is unproven at scale, and a reported mega-funding round has already stalled once. But it is also a coherent answer to a question the frontier labs have mostly waved away: what happens when models are good enough for most work, and the edge shifts from having the single best system to being able to make any system yours. For everyone tracking the AI build-out, and readers following it on Blumefield, Inkling is the clearest sign yet that the fight is moving from the model to the workshop around it. Thinking Machines has planted its flag firmly in the second camp, and it is betting the company that the rest of the market follows.