Models
Thinking Machines Releases Inkling As Open-Weight Model Competition Widens
Mira Murati's Thinking Machines Lab released Inkling, a 975-billion-parameter mixture-of-experts model with open weights, putting customization and developer control at the center of the next model race.
By Michael G ยท
Thinking Machines Lab has released Inkling, its first open-weight foundation model, turning Mira Murati's company from a heavily financed frontier-lab story into a model provider with a concrete technical stake in the market. The company's July 15 announcement describes Inkling as a 975-billion-parameter mixture-of-experts transformer with 41 billion active parameters, a context window of up to 1 million tokens and native reasoning over text, images and audio. Reuters and TechCrunch both framed the release as an attempt to offer developers a customizable alternative to closed models and to Chinese open-weight systems that have gained traction in 2026.
The model's design reflects a pragmatic tradeoff. A mixture-of-experts system can advertise very large total scale while activating a smaller portion of the model for a given request, which can reduce serving cost and latency. Thinking Machines says Inkling is not the strongest overall model available today. That admission is strategically useful. The company is arguing that raw leaderboard dominance is not the only path to adoption. A model that can be downloaded, customized and fine-tuned may be more valuable for teams that need control over behavior, data and cost.
The release also pairs the model with Tinker, Thinking Machines' fine-tuning platform. That pairing matters because open weights alone do not solve the operational problem. Most developers do not want to build training infrastructure from scratch, manage specialized hardware or design evaluation pipelines for every experiment. A model release becomes more useful when it is connected to a customization workflow that helps teams adapt the base model without turning every product group into a frontier-lab operations team.
Inkling's model card says the system is intended for English and general multilingual use, coding, retrieval-augmented generation, tool-use systems and general conversational applications. It is released under Apache 2.0, a permissive license that can matter for commercial developers. The model accepts text, image and audio inputs and produces text outputs, including code and structured data. Those choices put it in the center of the application layer where companies are trying to build agentic products.
The competitive context is important. Open models have become a geopolitical and commercial pressure point. Chinese labs have shown that open-weight releases can pull developers toward cheaper deployment paths, while U.S. frontier labs have leaned heavily on closed hosted access. Thinking Machines is entering that gap with a U.S.-based open-weight model and a commercial training service. That gives companies another option if they want more control than a closed API provides but more support than a raw model download offers.
The safety question is unavoidable. Open-weight models allow inspection and adaptation, but they also distribute capability beyond the original provider's direct control. Thinking Machines says Inkling has undergone safety testing, and its model card describes intended uses and acceptable-use limits. The harder test will come when downstream developers fine-tune the model for specialized tasks, wrap it in tools and deploy it in environments Thinking Machines cannot fully observe.
The market question is whether customization becomes a stronger selling point than general intelligence. Enterprises often do not need a model to win every public benchmark. They need predictable performance on their domain, transparent cost, secure deployment paths and a way to measure whether the model improved a workflow. If Inkling performs well enough and Tinker makes post-training accessible, Thinking Machines can compete on fit rather than fame.
Inkling is therefore less a single release than a statement about where model competition is moving. The first AI boom rewarded the company with the most compelling general assistant. The next one may reward the company that gives developers a strong base model, a controllable training loop and enough openness to make the model feel like infrastructure rather than a rented black box.
Topics: Thinking Machines, Inkling, open weights, models