Inkling-Small-6layer

A 6-layer slice of thinkingmachines/Inkling-Small, made for miles CI tests. It is not meant for inference quality: the layers were cut without any retraining, so its outputs are not meaningful.

• Text layers 0-5 of the original 42: five local (sliding-window) attention layers (0-4) followed by one global attention layer (5), the same pattern the full model repeats every six layers. Layers 0-1 have dense MLPs and layers 2-5 are MoE.
• config.json is the original config with only text_config.num_hidden_layers = 6 and text_config.local_layer_ids = [0, 1, 2, 3, 4] changed.
• Every tensor is a byte-identical BF16 copy from the original checkpoint: the embeddings, final norm and unembedding, text layers 0-5, all MTP layers, and the vision and audio adapters. The tokenizer, chat template and processor files are copied unchanged.