Thinking Machines Releases Inkling: 975B MoE Open-Weight Model Supports One Million Tokens and Native Multimodality

Thinking Machines Lab has released Inkling, its first open-weight model pretrained in-house. The model is a Mixture-of-Experts Transformer with 975 billion total parameters and approximately 41 billion parameters active during each inference pass. It was pretrained on 45 trillion text, image, audio, and video tokens and supports a maximum context window of one million tokens. It accepts text, image, and audio inputs and produces text output, with its weights licensed under Apache 2.0. The smaller Inkling-Small has 276 billion total parameters and 12 billion active parameters, but it remains in preview and its weights have not yet been fully released.

What stands out most about the model is not any single leaderboard ranking, but the integration of controllable reasoning cost and post-training within the same interface. Users can adjust effort and fine-tune the model through Tinker. The team also demonstrated the model writing its own fine-tuning tasks within an agent framework, creating data and evaluations, and then loading the new weights. For training, the company says it completed more than 30 million rollouts using asynchronous reinforcement learning and trained the model on NVIDIA GB300 NVL72 systems.

The company acknowledges that Inkling is not the strongest model overall and that some results on the release page were obtained from checkpoints different from the publicly available weights, with tests conducted between June 30 and July 13. This means the benchmark figures should not be treated as guarantees for the downloadable version. Engineers should next evaluate its VRAM requirements, expert-routing efficiency, effective long-context recall, and the maturity of Transformers, vLLM, SGLang, and quantization backends for multi-node deployment.