The raw mtp.* tensors extracted from exolabs/qwen3.6-35b-a3b-2.06gb-mtp-fp4-mirror, bf16 — just the MTP block, none of the target weights. exo-nvidia's MLX speculative-decoding path (SpeculativeStrategy.kind == 'mtp') loads these and sanitises them at load time (the +1.0 norm shift, gate/up split and per-expert stack are applied by the runtime, not here).