Skip to content
← Speculative & Constrained Decoding / 07
hardNewNVIDIATogether AIFireworks AI

How are DeepSeek V3's multi-token prediction modules trained, and how does the engine turn that one MTP module into a speculative drafter?

DeepSeek V3 ships a drafter it trained alongside the model for free. The answer walks the MTP module's wiring and loss, prices a draft step against a 37.6B-active target, and reproduces the reported 1.8x from first principles.

Updated Oct 2026 · Worked answers on inference mechanisms, with calculations, implementation details, and follow-up questions.

DeepSeek V3 ships a drafter it trained alongside the model for free. The answer walks the MTP module's wiring and loss, prices a draft step against a 37.6B-active target, and reproduces the reported 1.8x from first principles.

an account raises the free count per topic
UP NEXT ON YOUR JOURNEY
Next in this trackWhen does n-gram or prompt-lookup speculation work well with no draft model at all? Walk me through a worked acceptance example.Next in this trackYou turned on speculative decoding in production. What do you measure to prove it helps, and what is the right baseline to compare against?Next in this trackHow does grammar-constrained decoding build the allowed-token mask at every step, and what does that cost per step at batch 128?
DISCUSSION · 0

Nobody has commented on this answer. Add a correction or another way to work it.