Author(s): Prachi rise

Originally published on Towards AI.

Deepseek-V3: Multi-Token Prediction — Part 3

This is the full series of Deepseek-V3 technical report, where i explain all the technical details in simpler words with code implementation and explanation.

The article introduces multi-token prediction (MTP) as an extension of conventional next-token autoregressive modeling, explaining how training with multiple future tokens provides stronger learning signals. It then describes DeepSeek-V3’s approach using sequential MTP modules, where predictions at one depth feed into the next depth. The piece breaks down the internal MTP components (shared embedding, shared output head, transformer block, and a projection matrix), outlines how hidden states and future token embeddings are combined through normalization and projection, and shows the step-by-step flow of MTP depth k (including selecting future token embeddings, updating hidden states via transformer blocks, and producing logits with the shared head). Finally, it provides a PyTorch-style code implementation and concludes with the key takeaways: MTP improves representation learning for future-token prediction, DeepSeek-V3 stacks sequential MTP modules, and the method can also be leveraged for speculative decoding during inference.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI