Inkling: a ~1T-parameter open multimodal model with 1M context
Thinking Machines released Inkling on Hugging Face, a ~1T-parameter decoder-only multimodal MoE (41B active) that takes image, text and audio, has a 1M context, and was trained on 45T tokens; a smaller Inkling-Small followed.
What happened
Thinking Machines Lab released Inkling on Hugging Face, with day-zero support in Transformers, SGLang, vLLM and llama.cpp. A later update added Inkling-Small.
The technical details
From the Hugging Face post:
- a decoder-only multimodal Mixture-of-Experts model with 975B total / 41B active parameters
- natively takes image, text and audio inputs
- 1M-token context
- trained on 45 trillion tokens of text, images, audio and video
- intended for reasoning across modalities and for domain adaptation via fine-tuning
- released in BF16 and a calibrated NVFP4 variant, with speculative MTP layers
- Inkling-Small: 276B total / 12B active, same architecture, with MXFP8 and NVFP4 weights
Why it matters
It joins the trend of trillion-parameter open-weight MoEs (Kimi K3 is summarised here in July 2026), released already quantised, because that’s the only practical way most people can run them.
#llm#multimodal#open-weights#mixture-of-experts
Sources
- Welcome Inkling by Thinking Machines — Hugging Face
This is a summary written from the linked public sources. Figures and numbers are as reported there.