Thinking Machines Lab Releases ‘Inkling,’ a 975 Billion-Parameter Open Model

Thinking Machines Lab released Inkling, a 975B-parameter open-weights multimodal MoE model with 41B active parameters and a one-million-token context window.

Model Inkling

Developer Thinking Machines Lab

Release 2026-07-15

Architecture Mixture-of-Experts · 975B total / 41B active

Context Up to 1 million tokens

Input Text · image · audio

Thinking Machines Lab’s First Open-Weight Model

Thinking Machines Lab has released its first self-trained open-weight model, ‘Inkling.’ It uses a Mixture-of-Experts architecture that activates 41 billion of its 975 billion total parameters during inference and supports contexts of up to 1 million tokens. Pretraining used 45 trillion tokens comprising text, images, audio, and video.

Inkling directly understands not only text but also images and audio, and can adjust its inference budget to suit the task. Thinking Machines Lab presents it as a general-purpose model capable of handling knowledge, mathematics, and science; agentic coding and tool use; instruction following; and prediction confidence in a balanced way.

The full weights are available on Hugging Face, along with NVFP4 checkpoints for NVIDIA Blackwell environments. It can be deployed through services including Together AI, Fireworks, Modal, Databricks, and Baseten, or fine-tuned for specific domains through Tinker.

The significance of this release is that users can do more than simply call the model: they can download the weights, modify them directly, and operate the model themselves. Thinking Machines Lab has positioned Inkling as the first in a family of models that will expand to multiple sizes.

🏷 Other posts for the same work →

Maru