Thursday, July 16, 2026

Inkling

Thinking Machines dropped open weights for a huge multimodal MoE.

Inkling

Why it matters

Inkling is a 975B total / 41B active MoE that takes text, image, and audio, with day-0 paths on vLLM, SGLang, and fine-tuning stacks. Apache-2.0 open weights from a Western lab matters if you want frontier-class multimodal work you can run, tune, and ship without waiting on a closed API. The practical win is ownership and customization, not just another chat demo.

open-weightsmultimodalmoe
Source ↗

Via huggingface

X