Mira Murati threw a compact inverter of Inkling 2 Mini in 26.4B under Apache 2.0 for 16GB VRAM Thinking Machines Lab put the open weights of Inkling 2 Mini (26.4B) under the Apache 2.0 license - a junior model of a new line of loop variables that fits 16GB of memory of your f****** laptop and immediately got 0-0 support in vLL,Mthfth. The MoE architecture uses 12 common base layers through which data is operated in 3 passes with gateway messages passed between lectures. This allows a compact model to score 20.1 points on the artificial analysis benchmark and 48.2% on the DeepSWE V1.1 benchmark. At the end: • Architecture: loop MoE-transformer (26.4B parameters, 12 common layers x 3 sick), Apache license 2.0 - Day-0 to build vLLM on NVIDIA GB200 TP1-decoding Delay reduced from 10.29 to 3.84 ms (batch 1) and from 16.41 to 784 ms (batch 32) Hugging is already available official weight and Jacks) From Jackth HDL.