DeepSeek has made its V4 models answer up to 85 percent faster for each user, and it did not add a single GPU to do it. The speedup comes from DSpark, an inference technique the Chinese lab described in a research paper posted to arXiv in early July and has already switched on in its production serving system. The code and the trained checkpoints are free to download.
Giant new models grabbed most of this month’s attention. This release points the other way. Instead of making the model smarter, DeepSeek made the act of generating text cheaper, and for anyone paying for GPUs by the hour, that may be the more valuable kind of progress.
The waiting problem
Language models write one token at a time. Each new token requires a full pass through the network, conditioned on everything written so far, which means the time to finish an answer grows with its length. That was tolerable when chatbots produced a paragraph. It becomes painful now that reasoning models think out loud for thousands of tokens and agents chain long tasks together for minutes at a stretch. The better models get at thinking, the more time they spend stuck in this queue.
Speculative decoding is the standard escape hatch. A small, fast draft model guesses the next several tokens, and the big model checks the whole guess in one pass instead of generating each token itself. Checking is far cheaper than writing, and the acceptance rule guarantees the final output is identical to what the big model would have produced on its own. Nothing about the answer changes. It just arrives sooner.
Guess in parallel, check with judgment
The paper, titled “DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation,” attacks the two places where this scheme usually breaks down.
The first is draft quality. Drafters that propose tokens one by one are accurate but slow. Drafters that propose a whole block in a single pass are fast but grow incoherent toward the end of the block, because each guessed token cannot see the guesses before it. DeepSeek’s answer is a hybrid the paper calls semi-autoregressive: a heavy parallel backbone proposes the block at once, then a lightweight sequential head nudges each position based on the one before it. In offline tests across Qwen3 target models from 4 to 14 billion parameters, that design stretched the accepted draft length by roughly 27 to 31 percent over Eagle3, the leading autoregressive baseline, and by 16 to 18 percent over the parallel method DFlash.
The second problem is deciding how much of a draft is worth checking. A confidence head scores every proposed token on its odds of surviving verification. A scheduler then looks at how busy the serving system actually is. When GPUs sit idle, checking a long speculative tail costs almost nothing. When the system is packed, that same tail steals capacity from other users, so DSpark verifies only the prefix it believes in and discards the rest. Verification stops being a fixed habit and becomes a decision made per request, under live load.
The production numbers
Deployed behind real user traffic, the gains are large. Compared with MTP-1, the speculative decoding baseline DeepSeek previously ran in production, DSpark speeds up per-user generation by 60 to 85 percent on V4-Flash and by 57 to 78 percent on V4-Pro at the same overall throughput. The starker result shows up under strict speed guarantees. When the system must keep every user above 120 tokens per second on Flash, or 50 on Pro, the old baseline’s capacity collapses while DSpark keeps serving. DeepSeek says this unlocked interactivity tiers it simply could not offer before.
Run the arithmetic from the other direction and the story is about money. A fleet that handles the same traffic with markedly fewer GPUs, or returns faster answers from the same hardware, is a direct cut to the largest operating cost in the business.
An unusually open release
DeepSeek published the trained DSpark checkpoints for its V4-Flash and V4-Pro preview models on Hugging Face, and released DeepSpec, a training and evaluation codebase for speculative decoding that also implements Eagle3 and DFlash, on GitHub. That continues a pattern. The lab has been shipping serving infrastructure alongside its models for well over a year, material that mostly benefits teams operating large clusters rather than hobbyists, but that hands every open-weight model host a recipe to study.
None of this transfers automatically. Every model has different specs and every serving stack has different constraints, so the numbers above belong to DeepSeek’s system rather than to the technique in the abstract. The thing to watch is how quickly these ideas, particularly load-aware verification, surface in the open serving engines the rest of the ecosystem runs on. Inference efficiency rarely makes headlines the way new models do. It quietly sets the price of everything those models are used for.
For more coverage of AI research and infrastructure, visit Mylistingo.
Source: this article draws on the video “DSpark: DeepSeek-V4’s Insane Compute Optimization Explained” by the YouTube channel bycloud, alongside DeepSeek’s DSpark paper.
