Thinking Machines Lab ships Inkling 2 Mini: looped MoE with 3B active params for local setups

ariG23498 · x · 2026-11-21

Thinking Machines Lab released Inkling 2 Mini on Hugging Face, a looped-transformer multimodal MoE with 26B total and 3B active parameters per token. Some layers loop three times to boost effective capacity; it supports 1M-token context and image+text input. Ships in mixed precision with NVFP4/MXFP8 expert weights plus a DSpark drafter for speculative decoding, and runs locally via transformers, llama.cpp, SGLang, and MLX.

Original post →

More from Models

Models channel →