Former OpenAI CTO's first AI large model: architecture draws on DeepSeek-V3, improved with Kimi K2.5

```

Thinking Machines Lab, an AI startup founded by former OpenAI CTO Mira Murati, has released its first general-purpose AI model, whose architecture directly draws on the technological achievements of Chinese AI labs.

On July 15, according to the Financial Times, this foundational model, named Inkling, references DeepSeek’s DeepSeek-V3 at the architectural level and is optimized in the post-training stage using data generated by Moonshot AI’s Kimi K2.5. Inkling will be released as an open-weight model, allowing users to deploy it on their own servers and fine-tune it for specific scenarios.

The release comes as Chinese AI models are accelerating their catch-up in the global market. Data from OpenRouter shows that the overall usage of Chinese AI models has surpassed that of their US counterparts this year.

Inkling’s Positioning: Open Weights, Emphasizing Customization

According to benchmark data shared by Thinking Machines, Inkling’s overall performance is currently below Anthropic’s Claude and OpenAI’s flagship products, and also lags behind some leading Chinese models. Nevertheless, the company chose to release it in an open-weight format, offering differentiated competition compared to closed-source products like ChatGPT and Claude.

Open weights mean that enterprises and developers can deploy the model on their own servers, and conduct deep customization and fine-tuning according to specific business needs. Thinking Machines stated in their blog that most current AI tools are “fixed after training in a few places” rather than being shaped by users. “To extend human will and judgment, we need AI to be as diverse and distributed as humans themselves,” claimed the company.

The company has also launched the Tinker platform, specifically for enterprise clients, allowing them to fine-tune and customize large language models for specific business application scenarios. Notably, competitors like OpenAI also provide customizable open-weight models, and cloud platforms such as Amazon and Microsoft offer model fine-tuning services based on users’ own datasets. Thinking Machines faces significant competitive pressure in this field.

Technical Path: Leveraging the Chinese AI Ecosystem

Inkling’s borrowing of Chinese AI achievements in its technical construction is one of the most eye-catching details of this release. Since its launch earlier this year, DeepSeek-V3 has sparked widespread discussion in the global AI community due to its low cost and high performance, and its architectural design has become a blueprint for multiple emerging models. Kimi K2.5, launched by Moonshot AI, is an inference-enhanced model used in Inkling’s post-training optimization stage.

Thinking Machines stated that Inkling is a “from-scratch trained” model and provides complete weights for users. “Today, by releasing a model trained from scratch by us, with full weights, we are advancing our mission to allow people to make it their own,” the company said.

Thinking Machines' high profile in the market is largely attributable to Murati’s unique status in the AI industry. She previously led the development of several core OpenAI products, including ChatGPT, the image generation tool DALL-E, and voice mode capabilities, and played a key role in the temporary removal of CEO Sam Altman by OpenAI’s board in November 2023. Murati left OpenAI in September 2024 and founded Thinking Machines in February last year.

About a year after its founding, the company completed a $2 billion seed round, reaching a post-investment valuation of $12 billion. Investors include venture capital firm Andreessen Horowitz, chipmakers Nvidia and AMD, and hedge fund Jane Street.

In addition to the Inkling foundational model, Thinking Machines is actively expanding its product line and seeking to build a competitive advantage through differentiation. In May this year, the company previewed its first “interactive model,” which can perform actions based on user audio and video input, not limited to text understanding, signaling strategic intent to expand into multimodal and everyday life scenarios.

Risk Warning and DisclaimerThe market involves risks, and investment should be cautious. This article does not constitute personal investment advice and does not take into account the special investment goals, financial situation, or needs of individual users. Users should consider whether any opinions, views, or conclusions in this article suit their particular situation. Invest accordingly at your own risk. ```