DeepSeek Launches DSpark to Boost Inference Speed by 60% to 85%

DeepSeek released DSpark on June 27, a speculative decoding framework that accelerates per-user generation speeds by 60% to 85% on its DeepSeek-V4 Flash model and 57% to 78% on the Pro variant.

DSpark isn’t a new model. It’s an engineering optimization layered on top of existing DeepSeek-V4 checkpoints. The company didn’t need to train a bigger model to get meaningfully better performance.

DSpark uses what DeepSeek calls a “semi-parallel” method that combines high-throughput parallel generation with adaptive verification. Instead of generating and checking one token at a time, DSpark speculatively generates multiple candidate tokens simultaneously, then selectively verifies only the promising guesses.

The throughput gains are even more dramatic than the per-user speed numbers suggest. Depending on concurrency levels, DeepSeek reports throughput improvements ranging from 51% to 400%.

DSpark has already been deployed in live traffic, not just benchmarked in a lab. DeepSeek says it outperforms prior acceleration methods including Eagle-3 and DFlash.

DeepSeek open-sourced the accompanying training and evaluation codebase, called DeepSpec, alongside the DSpark research paper (arxiv:2606.19348). The DeepSeek-V4-Pro-DSpark model checkpoint is available on Hugging Face, and inference examples have been published on GitHub.

DeepSeek has tested the framework on open models including Gemma and Qwen, suggesting the optimization technique could have applications beyond DeepSeek’s own ecosystem.

DeepSeek was founded in July 2023 by Liang Wenfeng and is backed by High-Flyer, a Chinese quantitative hedge fund.

Decentralized compute networks like Akash, Render, and io.net are betting on a future where AI inference is distributed across permissionless hardware. The economics of those networks depend heavily on how efficiently models can run. A framework like DSpark, which delivers the same output quality at 60% to 85% faster speeds, changes the cost calculus for anyone running inference workloads on centralized clouds or decentralized GPU networks.

If a decentralized compute provider can serve 51% to 400% more requests with the same hardware, the unit economics of renting out GPU time shift dramatically.