Thinking Machines Lab releases Inkling, 975B parameter open-source AI model
In brief
- Thinking Machines Lab released Inkling on July 15, a 975B parameter model under Apache 2.0 license
- Inkling processes text, images, audio, and video with a 1 million token context window
- Mixture-of-Experts architecture activates only 41 billion parameters at any one time
- Company raised $2 billion at $12 billion valuation since February 2025
Model Architecture and Capabilities
Inkling packs 975 billion parameters but uses a Mixture-of-Experts design where only 41 billion parameters activate at inference time. This approach balances scale with computational efficiency. The model processes text, images, audio, and video simultaneously and supports a context window of 1 million tokens—useful for long-form reasoning and document analysis.
Training consumed 45 trillion tokens on NVIDIA GB300 NVL72 hardware built from scratch. Thinking Machines Lab also released Tinker, a customization platform designed to let researchers and developers fine-tune Inkling for specific tasks.
Open Release and Implications
"Anyone can download the weights from Hugging Face and do essentially whatever they want with them," according to the release materials. The Apache 2.0 license removes restrictions on commercial use and modification, a departure from many frontier labs' approach to model releases.
The company itself has no crypto component. Inkling carries no token, no blockchain integration. Yet the open release carries implications for decentralized compute networks. Decentralized compute marketplaces like Akash, Render, and io.net aggregate GPU resources for inference—infrastructure that stands to benefit from demand to run large models outside centralized cloud providers.
Company Status and Timeline
Thinking Machines Lab was founded in February 2025. Since then, the company raised $2 billion in funding at a $12 billion valuation. Reports suggest ongoing discussions could push valuations toward $50 billion. The lab faced staff departures in January 2026, though the Inkling release signals operational continuity. Shipping a model of this scale roughly 18 months from founding represents an aggressive pace by any standard.
