Mira Murati's Thinking Machines Lab launches first open-weight AI model, 'Inkling'

Mira Murati's Thinking Machines Lab launches Inkling, an open-weight AI model designed for enterprise fine-tuning

Thinking Machines Lab, led by former OpenAI technology chief Mira Murati, released Inkling on Wednesday as its first AI model, choosing an open-weight architecture that lets customers modify it with their own data. The move signals a deliberate strategy: compete not on raw frontier capability, but on customizability and control.

Inkling has 975 billion total parameters, with 41 billion active at any moment—larger than most open-source models but far smaller than closed frontier systems. It uses a mixture-of-experts architecture. Murati framed the model explicitly as a foundation layer for adaptation rather than a standalone powerhouse: "We trained it to be a broad, balanced foundation model strong across many domains, flexible enough to adapt. Inkling is not the strongest overall model available today, open or closed." That messaging differs sharply from typical model launches, where companies claim superiority in at least one domain.

Strategic fit with Tinker API

The open-weight decision makes structural sense given TML's core business. The company's Tinker API handles fine-tuning for customers; releasing open weights reinforces a Red Hat-style value prop: customers can fork the model if they choose, but TML becomes more valuable precisely because it helps them customize and maintain it. That integration and ongoing support justifies commercial fees even with open weights available.

Distillation shadows the claim

Jack Morris, founder of Engram Labs, claimed Inkling was "the only open weight model trained without distilling from OpenAI or Anthropic," contrasting it against Kimi, GLM, Qwen, and Nemotron. That claim drew immediate pushback. In TML's own blog post, the company states it "ran an initial supervised fine-tuning on synthetic data generated by open weight models including Kimi K2.5"—a practice broadly classified as distillation. Morris's team later issued a clarification apologizing for "unfounded claims" while maintaining that Inkling remains a strong open-source model.

The disagreement hinges on what counts as distillation. TML used synthetic data from other models for one piece of the post-training pipeline, not the entire training run. That's a lighter touch than Nvidia's Nemotron approach, which distilled more heavily. Whether that distinction matters depends on whether those distilled capabilities proved essential to Inkling's performance or were merely expedient shortcuts.

Timing against geopolitical headwinds

The launch arrives as Beijing signals intent to restrict Chinese AI model exports. A week prior, reporting indicated the Chinese government is eyeing curbs on overseas access to domestic models. Simultaneously, Anthropic's head of national security policy, Tarun Chabra, accused Chinese labs of systematic distillation campaigns at the Aspen Security Forum—naming ZeePu for GLM 5.2, and singling out DeepSeek for ongoing "adversarial" distillation. Anthropic now shuts down distillation accounts at a scale of millions per week, suggesting the problem is distributed and difficult to contain through conventional enforcement.

A Western, US-based open-weight model sidesteps both geopolitical risk and trust friction that US enterprises feel toward Chinese open-source models—regardless of whether they harbor actual security concerns or simply prefer domestic alignment.