On June 27, DeepSeek V4 quietly updated a framework called DSpark. The technical term is "an 80% increase in inference speed," which in plain language means—the speed at which the AI replies to your messages is nearly twice as fast.

First, let's explain why it was slow before.

For large language models to generate text, it's essentially a "word-guessing game." For every word the AI writes, it has to re-read and re-calculate all the previous words to decide what the next word should be. Writing 100 words means digesting what it has written 99 times. Academically, this is called "autoregressive generation"—the next word can't move until the previous one is calculated.

So every time you ask the AI a question, it replies word by word, like squeezing toothpaste—it's not intentionally slow; the technical principle dictates it can only work this way.

So how does DSpark solve this?

Here's an analogy to help you understand:

Before, it was like a senior professor personally writing word by word—high quality, but painfully slow.

DSpark's approach is: first, let a fast but average-skilled "assistant" draft, guessing several words ahead in one go, then present them all at once to the senior professor. The professor glances over it, directly keeps the correctly guessed words at the beginning, and starts writing a correct one from the first mistake.

This is called "speculative decoding"—draft first, then review quickly.

But there's a catch here: the faster the assistant guesses, the more likely it is to guess blindly later on. The paper calls this "suffix degradation"—the first word is okay, but by the fifth or sixth, it's basically making things up.

The brilliance of DSpark lies in this: it assigns a "reliability score" to each guessed word, then dynamically decides how many words to send to the senior professor for verification based on how busy the system is. When the system is idle, verify more; when busy, only verify the most reliable batch.

What's the result?

For responses of the same quality, speed is increased by 60% to 85%. A reply that used to take 10 seconds now comes out in five or six seconds. Most crucially—during peak hours, DeepSeek finally stops constantly "spinning."

Why am I quite moved?

The focus of this update isn't that the model got smarter, but that the engineering implementation was done beautifully. Without changing the model itself, it boosted speed purely by optimizing inference efficiency.

Behind this is the result of joint R&D by Peking University and DeepSeek, with the paper personally authored by Liang Wenfeng. Moreover, the entire technology is open-sourced—other AI companies can take it and use it.

Not burning money to pile on parameters, but relying on technology to squeeze out efficiency. This kind of pragmatic effort is worthy of respect.