Thinking Machines Lab launched Inkling on July 15, 2026, a 975-billion-parameter open-weight Mixture-of-Experts foundation model and the company’s first product-grade release. Founded by former OpenAI CTO Mira Murati, Thinking Machines Lab positioned Inkling as an open-weight alternative to proprietary systems, designed primarily as a starting point for organizational fine-tuning rather than as a general-purpose standalone model.
Thinking Machines Lab bets on open-weight AI, giving organizations a powerful, customizable foundation rather than another closed system.
Inkling’s architecture is a sparse MoE transformer with 975 billion total parameters but only approximately 41 billion active parameters per token, enabling lower inference costs and faster throughput at extreme scale. The model uses a 66-layer transformer stack with per-token routing to 6 of 256 experts, plus 2 shared experts operating in a shared hidden space. A hybrid local and global attention scheme supports long-context efficiency and stability across extended workloads.
The model was pretrained on approximately 45 trillion tokens spanning text, images, audio, and video, giving it native multimodal reasoning across text, image, and audio streams within a unified representation space. Despite the multimodal pretraining, the current public release is limited to text outputs, covering natural language, code, styled artifacts, and structured data. Support is advertised for English, additional languages, and multiple programming languages.
Inkling supports context windows of up to approximately one million tokens in long-context configurations, made practical by its hybrid local-global attention design, which avoids prohibitive memory growth at scale. Hosted APIs expose shorter windows, such as 256,000 tokens. The model also includes speculative Multi-Token Prediction layers for faster generation and an adjustable thinking effort knob, allowing users to trade off speed, cost, and answer quality depending on the workload.
On benchmarks, Inkling reported 87.9% on GPQA Diamond, a measure of graduate-level question answering difficulty, and 77.6% on SWE-bench Verified, which evaluates end-to-end software engineering task resolution. External evaluations described a clear performance step-up versus Nimotron 3 Ultra across broad benchmark suites.
Thinking Machines Lab entered a competitive landscape already populated by major closed and open model providers. By releasing Inkling as an open-weight model, the company signaled a strategic bet that organizations would prefer a highly capable, customizable base model over closed, one-size-fits-all systems.
The architectural choices, particularly sparse activation and the MoE routing design, reflect an emphasis on deployment practicality alongside raw capability, making large-scale inference more accessible to organizations without hyperscale infrastructure.
Inkling’s combination of a massive parameter count, multimodal pretraining, long-context support, and open-weight availability represents a significant technical statement from Thinking Machines Lab. Whether it translates into broad adoption will depend on how effectively organizations can fine-tune and deploy the model within their own operational constraints, which appears to be precisely the use case Thinking Machines Lab built it for. Fine-tuning is facilitated through the Tinker platform, where users can access the Inkling Playground for interactive experimentation and customization across diverse workflows.
