Inkling

• 🏢 Thinking Machines Lab's first open-weights model.
• 🧠 Sparse Mixture-of-Experts: 975B total, 41B active parameters.
• 📏 Up to 1M-token context window.
• 👁️ Natively accepts text, image, and audio inputs; outputs text.
• 🔧 66 layers with hybrid local/global attention.
• ⚡ Controllable "thinking effort" balances speed against reasoning depth.
• 🎯 Built for chat, coding, tool use, and agentic workflows.
• 🔒 Apache 2.0 license.

thinkingmachines

Thinking Machines is an AI research organization whose work centers on large language models. Its focus lies in building capable text-based systems, contributing to the growing field of generative AI.

Inkling is Thinking Machines Lab's debut model — released with open weights under an Apache 2.0 license. Because it is a first-generation model, it has no same-family predecessors to compare against; the lab positions it as a broad, general-purpose multimodal generalist meant to serve as a customizable base for a range of tasks.

Architecturally, Inkling is a 66-layer sparse Mixture-of-Experts model. Routing activates only a fraction of its experts per token, giving 975B total parameters but roughly 41B active per token, which keeps inference more efficient than a comparably sized dense model. Attention alternates between local and global layers, and a quantized NVFP4 variant is also published.

Multimodality is native: the model accepts text, image, and audio inputs and generates text, with all modalities processed jointly by the decoder. On Venice, video input is not supported, per the catalog description. The 1M-token context window suits long documents, extended conversations, and agentic workflows.

A defining feature is variable, controllable thinking effort, letting developers trade latency for reasoning depth depending on task complexity. The model is aimed at chat, coding, tool use, and function calling, and the lab frames it as a foundation intended for domain adaptation and fine-tuning rather than a narrowly optimized specialist.