Particle.news

Thinking Machines Releases Inkling, an Open‑Weights 975B Multimodal Model

The launch enables enterprise download plus fine‑tuning of Inkling through partner platforms, reducing run costs, retaining model control, and hosting models on internal systems.

Overview

  • Thinking Machines publicly released Inkling on Wednesday, July 15, 2026, as its first general‑purpose open‑weights model.
  • Inkling is a mixture‑of‑experts system with 975 billion total parameters and roughly 41 billion active parameters for any given task.
  • The model processes text, images, audio and video, supports very long context windows (reports cite up to 1 million tokens), and offers a controllable ‘thinking effort’ setting to trade compute for performance.
  • Company benchmarks and partner demos show competitive performance on agentic reasoning and coding tasks but not top overall results, and the lab says Inkling was trained on Nvidia GB300 systems with post‑training reinforcement learning and safety tuning.
  • Weights are available through Thinking Machines’ Tinker platform and partners including Databricks and Hugging Face, a move designed to push enterprises toward self‑hosting and customization while leaving independent benchmarking and long‑term monetization as open questions.