Thinking Machines, the San Francisco-based AI business founded by former OpenAI Chief Technology Officer Mira Murati, has launched its first general-purpose AI model, Inkling, an open-weight system with 975 billion parameters. Murati announced the debut on X on July 15, 2026. Open-weight means that the underlying model weights are freely available for anybody to download, run, and fine-tune, as opposed to proprietary closed-source models from OpenAI, Anthropic, or Google, which just allow access to API calls and do not reveal the model internals. Inkling is now available on Tinker, Thinking Machines’ own enterprise customization platform that launched in October 2025, as well as other developer platforms. Access is now free of charge.
The 975 billion parameter count makes Inkling one of the largest open-weight models released to date. Its architecture uses a Mixture-of-Experts design, which activates only 41 billion parameters at a time for any given task making it more computationally efficient than a dense model of the same total size despite the headline parameter count. The model is multimodal at input level, processing text, images, audio, and video, though it currently limits outputs to text and code.
Thinking Machines was explicit about what the model is and is not: company materials state directly that Inkling does not outperform every competitor on standard leaderboards, and that it is not the strongest model available among either closed or open systems. Instead, the focus was on customisability, lower hosting costs for companies that want to run models on private servers, and fine-tuning flexibility for enterprise clients.
“Our first model, Inkling. Trained from scratch, weights are open, fine-tunable on Tinker today.”~Mira Murati
Strong on Agent Tasks, Behind on Top Benchmarks:
Thinking Machines presented a full set of benchmark findings, comparing Inkling to closed models from Anthropic, Google, and OpenAI, as well as leading open solutions, the majority of which came from Chinese laboratories such as DeepSeek and Alibaba’s Qwen. Inkling trails the flagship closed models, like as OpenAI’s GPT-5.6 and Anthropic’s Claude Fable 5, on most broad benchmarks.
On coding jobs, Inkling is said to match Nvidia’s Nemotron 3 Ultra while utilizing roughly one-third of the tokens, showing a significant efficiency gain for that particular use case. According to Reuters, Inkling performed best on agent-related tasks, which require AI systems to perform a series of activities rather than answer single questions. That posture is deliberate: agentic AI is where industry demand is growing the fastest, and a customizable open-weight model that performs well on agent benchmarks fills a real market gap.
The company was founded with a team that was nearly two-thirds former OpenAI employees at launch. It was valued at $12 billion following an early-stage funding round led by Andreessen Horowitz. Thinking Machines has stated its goal is to build AI systems that are safer, more reliable, and targeted at a broader range of applications than those offered by its rivals.
“AI startup Thinking Machines revealed a new AI model that could serve as one of the few alternatives to popular open-source offerings from Chinese AI labs. Named Inkling, the 975 billion parameter model is open-weight. It is the first general-purpose model from the startup founded by former OpenAI CTO Mira Murati.”~Reuters
Why an Open-Weight Model and Why Now?
The open-weight decision is both a philosophical and strategic one. Murati has spoken publicly about the importance of AI development being more transparent and accessible, and Thinking Machines has been explicit that it sees open-weight models as a meaningful alternative to the increasingly closed model ecosystem dominated by OpenAI, Anthropic, and Google. The release of Inkling positions Thinking Machines as a Western counterweight to the Chinese open-source models primarily from DeepSeek and Qwen that have come to dominate the open-weight space since late 2024. Reuters described Inkling as potentially serving as one of the few Western alternatives to those Chinese offerings, noting that the open-source market has tilted heavily toward models coming out of Chinese labs.
Enterprise clients who want the performance and flexibility of a frontier-scale model without routing sensitive data through a third-party API or paying the compute costs associated with running a dense 975 billion parameter model will benefit from the Mixture-of-Experts architecture, which activates only 41 billion parameters per task. Tinker was designed with those clients in mind, and Inkling provides a showcase model for the platform.
“Thinking Machines, the AI startup from former OpenAI CTO Mira Murati, just launched its first AI model. Called Inkling, the 975B parameter open-weight model is positioned as a Western alternative to Chinese open-source AI. It is available now on Tinker and other developer platforms.”~TechCrunch
From OpenAI’s CTO Chair to Her Own Frontier Model in Under Two Years:
Murati left OpenAI abruptly in September 2024, just weeks before the company completed a $6.6 billion funding round that valued it at $157 billion. She had been at OpenAI for six years, overseeing the launches of GPT-4, DALL-E, Codex, and the early rollout of ChatGPT. Her departure was widely covered and came alongside a broader wave of senior OpenAI exits.
Within months, she had formed Thinking Machines, built a team mostly comprised of former employees, achieved a $12 billion valuation, and introduced Tinker as the company’s first commercial product. Inkling’s introduction completes the second leg of that journey, which includes a flagship model to complement the platform that hosts it.
The open-weight approach, the honest benchmarking, the focus on enterprise customisation, and the clear-eyed acknowledgement that Inkling is not yet the best model available in any category all point to a company trying to build trust with technical users before competing head-on with OpenAI or Anthropic on raw performance. Whether that positioning works, and whether Thinking Machines can close the performance gap as it iterates, will determine how much of the open-weight market it can credibly claim from DeepSeek and Qwen.
“Thinking Machines Lab founded by former OpenAI CTO Mira Murati has launched Inkling, its first AI model. It has 975 billion parameters, open weights, and is available on Tinker. Murati: ‘Trained from scratch, weights are open, fine-tunable on Tinker today.'”~The Verge
