Magistral-Small-2509 config.json

model_type mistral3. text_config: hidden 5120, 40 layers, 32 q heads, 8 kv heads, head_dim 128, intermediate 32768, max_position_embeddings 131072, model_type mistral, no sliding window. vision_config: 24-layer 1024-hidden pixtral encoder. Note the text_config model_type is plain 'mistral', i.e. the architecture is Mistral Small's dense text tower rather than a new geometry. | SOURCE AUDIT 2026-10-03: kind was 'spec-sheet' but the URL is a raw Hugging Face config.json served as text/plain, not a vendor spec sheet; set to 'repo'. The schema field it fills (layer/window/head counts) is machine-readable model metadata and should be read as such.