Mira Murati, the former CTO of OpenAI, has finally stepped out of the shadows with her startup, Thinking Machines Lab. Her debut project is Inkling, a multimodal model with open weights. Boasting a massive 975 billion parameters, Inkling claims the title of the most powerful open-weights solution developed in the United States. Murati is clearly aiming for the gap between closed-source giants like GPT-4o and the rapidly advancing Chinese competitors. However, behind these impressive figures lies a dangerous paradox: while the model shows an outstanding ability to mimic workflows, it is a chronic liar.
Architecture and the Bet on Autonomous Agents
From a technical standpoint, Inkling utilizes a classic Mixture-of-Experts (MoE) transformer architecture. Out of nearly a trillion parameters, only 41 billion are active at any given time, a design intended to ensure acceptable speeds despite a massive one-million-token context window. According to data from Artificial Analysis, Inkling scored 41 on the Intelligence Index, surpassing other American open models such as Nemotron 3 Ultra and Gemma 4. The developers intentionally "sharpened" the system for agentic tasks—simulating complex chains of intellectual labor where the neural network doesn't just answer questions but actively manages tools.
"Inkling is not the strongest general-purpose model on the market," Thinking Machines Lab candidly admits, suggesting it should instead be used as a foundation for customization on their proprietary Tinker platform.
In GDPval-AA v2 tests, which focus specifically on agentic logic, Inkling achieved an Elo rating of 1238. For businesses trying to build autonomous assistants capable of executing multi-step instructions, this looks like a robust American alternative. Essentially, Murati is offering a tool for those ready to fine-tune a model for niche tasks without handing their data over to the clouds of Altman or Anthropic.
Hallucinations and the "Chinese Trace" in Economics
The excitement over Inkling’s agentic talents fades when confronted with its catastrophic factual accuracy. Experts at Artificial Analysis have recorded a hallucination rate of 63%. This isn't just a margin of error; it's a dealbreaker for the corporate sector, where the cost of a mistake in documentation or code is too high. In this arena, Inkling loses decisively to Chinese open-source leaders like DeepSeek and Kimi, which deliver far more stable results at a lower inference cost. For now, Inkling's economics look questionable: the model’s enormous weight requires colossal compute power that isn't justified by the accuracy of its outputs.
The situation is further complicated by the origin of its data. Thinking Machines Lab fed the model 45 trillion tokens but admitted to using the Chinese model Kimi K2.5 to generate synthetic training sets. Furthermore, Murati’s team openly stated that some public data used might be subject to copyright. This looks like a desperate attempt to catch a departing train: leveraging others' work and ignoring legal risks to prevent American open weights from falling permanently behind. For now, Inkling feels more like a bold experiment than a production-ready business solution, and a 63% fabrication rate is a steep price to pay for "agentic" logic.