Inkling Mini - 26B A3B Thinking Machines released Inkling-2-Mini, a 26B-parameter image-text-to-text model on Hugging Face that the release claims matches or beats Qwen 3.8 27B on most benchmarks. The model uses a looping architecture that loops 3 times per token. The release gives no further benchmark figures or technical details beyond the parameter count and loop count. huggingface.co thinkingmachines/Inkling-2-Mini ยท Hugging Face Image-Text-to-Text model with 26B parameters. Download, explore files and run it on Hugging Face. As good or better than Qwen 3.8 27B on most benchmarks. Looping architecture - it loops a total of 3 times per token. I too am interested in seeing this optimized for 2xr9700s.