Adds the Multi-Token Prediction (MTP) head of Qwen3.8-Flash-Next to the tiny model, as in the real checkpoint.
What changes
model.safetensors: 31 mtp.* tensors (one QSA decoder layer with hyper-connections and MoE, fc_embedding, fc_hidden, the pre-fc norms and the head's hyper_connection_mixer), named as in the real checkpoint.
config.json: the real mtp section (1 QSA layer) and mtp_num_hidden_layers 0 → 1.
README.md: a model card describing the architecture, the initialization and the health checks.
Compatibility
The 243 main-model tensors are bit-identical to the current version, so existing tests are unaffected.
transformers ignores mtp.* when loading (_keys_to_ignore_on_load_unexpected).
Initialization and health checks
The MTP weights use the same init scheme as the main model: layers std 0.25, router std 1.
Checked on text, image and video prompts with a reference MTP forward that follows vLLM qwen4_exp:
logits are finite, with std ≈ 4 and median greedy top-2 margins of 0.55–0.82;
73–100% of positions get distinct predictions;
all experts are used, and its QSA drops 23–93% of the visible tokens;
the head depends on both of its inputs (next-token embedding and hidden streams);
