Thinking Machines Lab releases Inkling 2 Mini, a 26B local multimodal MoE model

·2026.11.20 09:00

Key point

The Looped Transformer-based model features 3B active parameters per token and outperforms Qwen3.8-27B on 32 of 48 benchmarks.

1 / 8

Details

  • Thinking Machines Lab released Inkling 2 Mini on November 20, 2026, a multimodal Mixture-of-Experts model optimized for local AI setups.
  • The model has 26B physical parameters with 3B active parameters per token and supports a 1M token context window.
  • It achieves State-of-the-Art (SOTA) performance in its class on the Artificial Analysis Intelligence Index v5, outperforming Qwen3.8-27B, Gemma 4 31B, and Nemotron 3.5 Super VL on 32 of 48 benchmarks.
  • Architecture utilizes a Looped Transformer design where 12 middle layers repeat 3 times, effectively creating 48 logical layers from 24 physical ones.
  • Supports mixed precision formats (NVFP4 and MXFP8) and includes a DSpark drafter for speculative decoding to improve speed.
  • Day 0 support is available for Transformers, Llama App (llama.cpp), MLX, and SGLang, with specific optimizations for NVIDIA B300 and Apple M4 Max hardware.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.