Skip to content
ViqusViqus
Navigate
Company
Blog
About Us
Contact
System Status
Enter Viqus Hub

Inkling 2 Mini: Looped MoE Model Claims Top Local Performance in Its Class

Mixture-of-Experts Looped Transformer Local LLM Inference Speculative Decoding NVFP4 Quantization GGUF Multimodal Models
November 20, 2026

This summary and analysis were generated by AI from the original article at Hugging Face Blog and may contain errors (how Viqus works). Read the source for full details.

Viqus Verdict Logo Viqus Verdict Logo 5
Clever Loops, Vendor-Judged Wins
Media Hype 4/10
Real Impact 5/10

Article Summary

Thinking Machines Lab has released Inkling 2 Mini, a multimodal mixture-of-experts model with about 26B total and 3B active parameters per token, designed for local setups. Its main architectural idea is a looped transformer: twelve middle layers run three times, so the model executes 48 logical layer visits using 24 physical layers. Each pass reads messages from the previous pass through learned gates. The model supports a 1M-token context window and image plus text input, and it ships in mixed precision with NVFP4 and MXFP8 expert weights. A DSpark drafter is included for speculative decoding. The company reports that Inkling 2 Mini leads 32 of 48 benchmarks against rivals including Gemma 4 31B and Qwen3.8-27B, and says it is state of the art in its class on the Artificial Analysis Intelligence Index v5. Its own vibe tests are informal, and the authors note that more validation is needed. Support is available through transformers, llama.cpp, MLX and SGLang, with GGUF quants provided. A research variant accepts a variable number of loop passes for faster inference.

Key Points

  • Inkling 2 Mini uses a looped transformer design that reuses middle layers three times, giving 48 logical layer visits from 24 physical layers.
  • The company reports leading 32 of 48 benchmarks against comparable open models, though the claims rely on its own selection and informal vibe tests.
  • The model is available across transformers, llama.cpp, MLX and SGLang, with GGUF quantizations and a speculative decoding drafter for faster local inference.

Why It Matters

This is a notable open-weight release for local deployment, and the looped-layer approach is a interesting architectural variation that could influence how small models trade parameters for compute. However, the performance claims come from the vendor's own benchmarks and limited vibe tests, and the article itself calls for further validation. Its impact on the industry is likely moderate and confined largely to the local-model community rather than a field-wide shift.


You might also be interested in