Murati Knows OpenAI’s Secrets. Her AI Suggests She Prefers China’s.

Thinking Machines, a new AI lab founded by former OpenAI CTO Mira Murati, has released its first model, Inkling. The model is an open-weight mixture-of-experts (MoE) transformer that utilizes a Chinese-inspired architecture and synthetic data from Moonshot AI’s Kimi models to compete in the enterprise AI market.

Thinking Machines Launches Inkling with $12 Billion Valuation

Thinking Machines launched Inkling this week after securing a record two billion dollar seed round. The company, founded in February 2025 by Mira Murati and a group of senior OpenAI researchers, entered the market with a twelve billion dollar valuation before shipping its first product. Financial backing for the venture includes Nvidia, AMD, Cisco, Andreessen Horowitz, and Jane Street.

Inkling is a mixture-of-experts transformer featuring 975 billion total parameters, with 41 billion active per token. According to company specifications, the model was pretrained on 45 trillion tokens across text, images, audio, and video, and supports a context window of one million tokens.

Inkling Performance vs. Chinese Open-Weight Models

Despite its scale, Thinking Machines admits Inkling is not the strongest model currently available. On industry benchmarks including Humanity’s Last Exam, Terminal Bench, and SWE-Bench Verified, Inkling trails leading Chinese open models such as Zhipu’s GLM 5.2 and Moonshot’s Kimi K2.6.

The technical architecture of Inkling closely follows the design of DeepSeek-V3. Furthermore, Thinking Machines confirmed that Inkling’s post-training process used supervised fine-tuning on synthetic data generated by open-weight models, specifically Kimi K2.5 from Moonshot AI.

Strategic Pivot to “Compliant” Enterprise AI

Thinking Machines is positioning Inkling as a low-risk alternative for American enterprises. While Chinese models like GLM and Kimi are often more capable and cheaper—accounting for roughly 45% of enterprise tokens routed through OpenRouter—they carry increasing regulatory risk in the U.S.

The U.S. State Department recently warned American companies about the risks associated with Chinese models. Simultaneously, House committees have launched probes into firms like Airbnb and Cursor over their use of Chinese AI. By releasing Inkling under an Apache 2.0 license and pairing it with “Tinker,” a proprietary fine-tuning platform, Thinking Machines provides a “compliant” base model for companies that cannot risk using foreign AI due to government procurement bans or congressional subpoenas.

Implications for OpenAI’s Technical Moat

The decision by Mira Murati—who previously oversaw OpenAI’s technology stack and served as interim CEO during the November 2023 board crisis—to build on Chinese blueprints suggests a shift in the AI frontier. The use of DeepSeek-style architecture and Kimi-generated data indicates that the most efficient paths to high-performance open research currently reside in Chinese labs.

This development coincides with a perceived deceleration at OpenAI. Kimi K2.6 recently surpassed GPT-5.5 on the SWE-Bench Pro benchmark, marking the first time an open-weight model beat a leading proprietary model on that specific metric. This compression of the gap between closed U.S. labs and open Chinese labs suggests that the proprietary “moat” once held by companies like OpenAI is shrinking.

Comparative Analysis: Open AI Ecosystems

The current trajectory suggests a structural disadvantage for U.S. firms if policy continues to wall them off from the highest-performing open inputs. While international competitors in hubs like Singapore or Barcelona can fine-tune the strongest available foundations, U.S. companies may be forced to build on weaker, “permitted” alternatives like Inkling.