Mira Murati's Thinking Machines Ships Inkling, a 975B-Parameter Open Model That Beats US Rivals but Trails China

Thinking Machines Lab has released its first open-weight language model, and the headline number is hard to miss: 975 billion total parameters. The model, called Inkling, marks the debut public product from the startup Mira Murati founded after leaving her post as OpenAI's CTO, and it arrives as the company's first real test of whether its research bets translate into a model that matters.

Inkling is built as a mixture-of-experts system, activating just 41 billion of its 975 billion parameters for any given task. That architecture keeps inference costs manageable while still allowing the model to draw on a much larger pool of learned knowledge. Training involved 45 trillion tokens spanning text, images, audio, and video, and the model natively processes all three input types, though its output today is limited to text, including code, structured data, and styled artifacts. It also supports a context window stretching up to one million tokens, putting it in the same league as the industry's longest-context systems.

On the Artificial Analysis Intelligence Index, Inkling scored 41, edging out Nvidia's Nemotron 3 Ultra (38) to become the top-ranked open-weights model from a US lab. That's a meaningful marker for Thinking Machines, but the scorecard isn't uniformly flattering: reported hallucination rates hit 63 percent against a factual accuracy score of just 40 percent, and by several accounts Inkling still trails the strongest open models coming out of Chinese labs. The company appears to be framing this release less as a frontier-beating flagship and more as a proof point for its stated thesis that one-size-fits-all AI is the wrong approach, following its earlier launch of Tinker, a platform for customizing models to specific tasks.

The release also puts fresh scrutiny on Thinking Machines' extraordinary capitalization. Founded in February 2025 by Murati alongside OpenAI cofounder John Schulman and former OpenAI VP of Safety and Robotics Lilian Weng, the company raised what's been described as the largest seed round in history, valuing it at $12 billion before it had shipped a single product. Inkling is the clearest signal yet of what that capital is buying: a team willing to publish an imperfect but competitive open model rather than sit exclusively on closed research.

For developers and enterprises evaluating open-weight options, Inkling's positioning is nuanced. Its scale and context length make it a serious contender for tasks that benefit from broad multimodal input and long-document reasoning, and its Intelligence Index ranking suggests real technical strength relative to domestic peers. But the elevated hallucination rate is a genuine caveat for any deployment requiring factual reliability, and teams will likely want guardrails or retrieval augmentation before trusting Inkling with unsupervised, fact-sensitive workloads. Whether Thinking Machines follows this release with output-side multimodality, image or audio generation, rather than just multimodal input, will be a key signal of how fast the company can close the gap with both US and Chinese frontier labs.