Thinking Machines launches Inkling, a 975-billion-parameter open-weights AI model
Inkling supports a 1 million token context window and is available for fine-tuning on the Tinker platform, alongside a preview of a smaller variant.
Thinking Machines has released Inkling, an open-weights multimodal Mixture-of-Experts model featuring 975 billion total parameters with 41 billion active parameters. The model supports a context window of up to 1 million tokens and was pretrained on 45 trillion tokens comprising text, images, audio, and video. A preview of a smaller variant, Inkling-Small, with 12 billion active parameters, was also shared.
The company, founded by Mira Murati, John Schulman, and Lilian Weng, secured the largest seed funding round in history, valuing the startup at $12 billion. Inkling is designed to be downloaded and modified by researchers and startups, contrasting with closed-model approaches. The model features controllable thinking effort to balance performance with token efficiency and is available for fine-tuning on the company’s Tinker platform.
Inkling utilises a sigmoid-based router with auxiliary-loss-free load-balancing bias and interleaves sliding-window and global attention layers at a 5:1 ratio. The model employs a hybrid optimisation strategy, using Muon for large matrix weights and Adam for other parameters. Post-training involved large-scale asynchronous reinforcement learning with over 30 million rollouts, resulting in log-linear improvements in reasoning performance.
The model was trained for epistemics, including calibration, instruction following, and resistance to censorship, with external safety testers verifying its behaviour on benchmarks such as FORTRESS and StrongREJECT. Inkling reportedly uses one-third the tokens of Nemotron 3 Ultra to achieve similar performance on Terminal Bench 2.1. Full weights are available on Hugging Face, including an NVFP4 checkpoint for efficient inference on NVIDIA Blackwell systems.
Thinking Machines stated that Inkling was trained on NVIDIA GB300 NVL72 systems. The model is accessible via APIs on Together, Fireworks, Modal, Databricks, and Baseten. The company offered a 50% discount on Inkling for a limited time on the Tinker platform.
