{
  "architecture": "wisp_recursive_shared_module",
  "hidden_state_stage": "trunk_pre_final_norm_residual",
  "mtp_depth_trained": 2,
  "mtp_layers": 1,
  "note": "One shared MTP module applied recursively, Qwen3-Next style. It consumes the trunk hidden state at position i and the embedding of the token at i+k, and predicts the token at i+k+1. The LM head is shared with the trunk, which ties both computations to one output projection but does not guarantee close distributions. The module contains a transformer block whose attention was trained under a causal mask over the whole window: at inference it must be given the sequence, not a single position.",
  "recursive": true,
  "requires_full_sequence_attention": true,
  "schema_version": 1,
  "shared_lm_head": true,
  "tensor_prefix": "mtp.",
  "trained_steps": 19073
}