Thinking Machines Lab releases Inkling, an open-weight model challenging one-size-fits-all AI
The AMW Read
Inkling is a new entrant in foundation models from a founder with top-tier credentials, updating the player map and advancing the open-weight distribution debate, though initial benchmarks do not claim frontier-level performance.
Thinking Machines Lab releases Inkling, an open-weight model challenging one-size-fits-all AI
Thinking Machines Lab, the AI startup founded by former OpenAI CTO Mira Murati, released its first in-house AI model, Inkling, on Wednesday. Inkling is an open-weight, mixture-of-experts system with 975 billion total parameters, activating roughly 41 billion per task. Trained on 45 trillion tokens across text, image, audio, and video, it natively reasons across all four modalities but currently outputs only text. The model is designed for customization via the company's Tinker platform, allowing organizations to fine-tune it for specific use cases. Inkling does not claim top-tier benchmark performance but emphasizes calibrated uncertainty flagging and adjustable thinking effort.
Why it matters: Inkling exemplifies the "open-weight distribution moat" pattern — the strategic bet that organizations will prefer customizable, self-hosted models over locked-in proprietary APIs from labs like OpenAI and Anthropic. This release updates the player map for foundation models (Segment 01), adding a new entrant backed by prominent founding talent and explicit anti-closed-model positioning. The model's debut also lends weight to the recurring "fastest-ARR-ramp" pattern, with Thinking Machines claiming it reached market in roughly nine months compared to three to five years for peers. Critically, the article reports that Inkling was partly trained on outputs from competitors' models via distillation, a practice that remains an open regulatory and ethical flashpoint across the industry.
The expert take: Inkling's release is a deliberate challenge to the "one-size-fits-all" model paradigm dominant among frontier labs. By offering open weights and a fine-tuning platform, Thinking Machines is betting on a future where enterprise AI differentiation comes from proprietary data and domain-specific calibration rather than generic model strength. The company's stated performance parity with Nvidia's Nemotron 3 Ultra at one-third the token count underscores the cost-efficiency argument. However, the model's reliance on distillation and its positioning as a starting point rather than a finished product means downstream safety and compliance responsibilities shift to customers, a model that may deter regulated industries. Inkling is less a GPT-killer than a proof point for a parallel ecosystem — one where control, not capability ceiling, defines competitive advantage.
#ThinkingMachinesLab #Inkling #OpenWeight #FoundationModels #AIEnterprise #ModelCustomization
