The JetBrains/Mellum2.1-12B-A2.5B-Thinking-MTP repository on Hugging Face highlights an architectural detail that shifts the economics of self-hosted inference. This is not a model built from scratch: it is a variant of Mellum2.1 Thinking with an added multi-token prediction (MTP) head. The head is an extra transformer block trained together with the main model, and it tries to predict the token after the one just sampled. Inference engines use it as a speculative draft: the model verifies every proposed token, so the output stays identical while decoding runs faster.

According to the repository card, the speedup is 1.4–1.8x over standard decoding, with about 90% of proposed tokens accepted on the first attempt. The head weights live in model-mtp.safetensors under the mtp. prefix, while config.json announces it with num_nextn_predict_layers: 1. Everything else is identical to Mellum2.1 Thinking.

The technical detail has a practical consequence that goes beyond a single checkpoint. Classic speculative decoding usually requires a separate draft model: a smaller LLM that proposes sequences which the main model then verifies. That approach increases VRAM usage and adds a second component to manage in serving. An integrated MTP head removes exactly that burden: the draft comes from the same checkpoint, with no extra model to load and synchronize. For teams operating on-premise, where GPU memory and operational complexity matter as much as throughput, the shift is significant.

This is not just about architecture. Support is already present in the three most common environments for local serving. vLLM from version 0.33 accepts --speculative-config '{"method": "mtp", "num_speculative_tokens": 1}'. transformers from version 5.20 exposes use_mtp=True in the generate method. llama.cpp, through the GGUF repository, uses --spec-type draft-mtp --spec-draft-n-max 1. In practice, a team that updates the runtime can enable the acceleration without changing the model and without building a parallel speculative decoding pipeline.

This is where the story becomes structural. The optimization moves from the orchestration layer to the checkpoint. The model carries an acceleration capability that frameworks only need to recognize. That reduces performance variability across installations and simplifies hardware sizing: teams gain speed on the same infrastructure without reserving memory for an external draft model. In TCO terms, the benefit does not come from adding GPUs, but from using the existing ones more efficiently. For teams evaluating self-hosted deployments, faster decoding without additional components lowers a concrete barrier.

It remains to be seen how the acceptance rate behaves on domains far from the base model's training data. However, token verification guarantees that output stays unchanged even when the draft is rejected: the risk is only a smaller speedup, not a quality degradation. The spread of MTP heads in publicly shared checkpoints signals that this technique is moving out of research literature and into everyday serving.