Particle.news

Thinking Machines Releases Inkling, an Open-Weights Multimodal MoE Model

Releasing open weights through Hugging Face, the sparse Mixture-of-Experts design aims to cut inference cost, raising fresh safety and misuse questions.

Overview

  • Thinking Machines Lab has publicly released Inkling and published its model weights on Hugging Face and the company API, making the full model available for download and use.
  • Inkling is a from‑scratch multimodal Mixture‑of‑Experts model with about 975 billion total parameters, roughly 41 billion active parameters per request, a 1 million‑token context window, and a reported pretraining corpus of around 45 trillion tokens.
  • The company positions Inkling to trade peak raw performance for lower runtime cost by activating only a small subset of experts at inference, a design meant to reduce latency and compute expense for users.
  • Independent benchmark summaries show mixed results: Inkling scores highly on some multimodal and math tests but trails top closed models and several leading open models on coding and advanced reasoning tasks, with GLM and DeepSeek often outperforming it in those areas.
  • Thinking Machines says it performed internal safety evaluations and offers the Tinker micro‑finetuning tool to let developers customize Inkling on modest hardware, but it acknowledges more work is needed to harden safeguards given the open‑weight release and the increased ease of customization.