Jin Daily AI Trivia - Deepseek release DSpark, increase 85% token speed

Jin Daily AI Trivia - Deepseek release DSpark, increase 85% token speed

There is a technique called speculative decoding that makes LLM output faster.

The idea is simple:

Instead of asking the big model to generate token by token, we let a smaller draft model or draft module guess a few future tokens first. Then the big model checks the whole draft in one shot.

Correct token? Keep it. Wrong token? Throw it away and let the big model take over.

So the speedup comes from the big model verifying multiple tokens at once, instead of generating one token at a time.

But the problem is, the longer the draft, the easier the later tokens go wrong.

DeepSeek’s DSpark improves this with a semi-autoregressive method.

It first uses a parallel draft backbone to generate a whole block fast, then adds a tiny Markov head to fix the relationship between nearby tokens. This keeps the GPU-friendly speed of parallel drafting, while reducing the quality drop at the end of the draft.

They also add a confidence scheduler. When GPU load is low, verify more tokens. When GPU load is high, verify fewer tokens to avoid wasting compute.

So DSpark is not really a new model. It is more like a production-grade turbocharger for DeepSeek-V4 inference.

Same brain. Less waiting. More token go brrrr.