Inkling 2 Mini, MLX conversion This work is derived from thinkingmachines/Inkling-2-Mini (revision 6e7c8248f75265bdd4fbab5273304243b5d8c2cb) by Thinking Machines Lab, Inc., released under the Apache License, Version 2.0 (see LICENSE). Use of the model is also governed by the Thinking Machines Model Acceptable Use Policy: https://thinkingmachines.ai/model-acceptable-use-policy/ Changes made here: the weights were converted to MLX format with MLX-VLM plus Inkling 2 Mini support (https://github.com/internet-projects/inkling2-mlx, v0.1.0). In the 4-8bit build the routed experts keep their released NVFP4 (layers 0-19) and MXFP8 (layers 20-23) values, repacked to MLX's nvfp4 / mxfp8 formats, and all other weights are unchanged. In the 3bit, 2.5bit and 4bit builds the routed experts were dequantized from those values and re-quantized to affine integer formats, and the attention, embedding, output and vision linear layers were quantized as described in the model card. The DSpark draft (mtp.safetensors), tokenizer, chat template and processor files are included unchanged. Conversion by NetProjects (Internet Projects Ltd), https://netprojects.com. Not affiliated with Thinking Machines Lab.