DeepSeek and Peking University Open Source DSpark to Cut LLM Inference Costs

Quick Facts

• DSpark, co-developed by DeepSeek and Peking University, boosts LLM inference speeds by 60% to 85% on DeepSeek-V4-Flash and 57% to 78% on DeepSeek-V4-Pro with no hardware changes required.
• The framework is open-sourced under the MIT license on GitHub and Hugging Face, alongside a companion toolkit called DeepSpec for training and evaluating draft models.
• Under strict latency conditions, DSpark delivered a 661% throughput gain on DeepSeek-V4-Flash at a 120 token-per-second service-level agreement.

DeepSeek and Peking University researchers released DSpark on June 27, 2026, an open-source speculative decoding framework that accelerates AI model inference by up to 85% without requiring new hardware or model retraining. The paper is formally titled DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation and was co-authored by DeepSeek founder Liang Wenfeng.

The release targets one of the most pressing cost pressures in enterprise AI deployment. Deloitte estimates inference will account for two-thirds of all AI compute spending, making efficiency gains at this layer directly relevant to operating budgets.

How It Works

Standard autoregressive models generate text one token at a time. For large architectures like DeepSeek-V4-Pro, a 1.6-trillion-parameter model with 49 billion active parameters, that sequential process creates severe GPU memory bandwidth bottlenecks.

DSpark addresses this with two technical advances. The first is semi-autoregressive generation, which retains the speed of parallel draft models while adding a lightweight sequential module to model dependencies between tokens within each block. The second is confidence-scheduled verification, which uses a confidence head to evaluate how likely each candidate token is to be accepted, then dynamically adjusts verification length based on real-time GPU load.

Unlike approaches that train a separate smaller draft model, DSpark grafts the speculative head directly onto the target model, cutting layer duplication. The result requires only 27% of the single-token inference FLOPs and 10% of the KV cache used by DeepSeek-V3.2 at a one-million-token context.

Benchmark Results

In head-to-head comparisons, DSpark showed 26.7% to 30.9% higher accepted token length than Eagle3, the strongest autoregressive baseline, and 16.3% to 18.4% higher than DFlash, the strongest parallel baseline.

Developer Rafael Caricio published real-world benchmarks in a GitHub pull request showing 26.33 tokens per second without speculative decoding, 39.88 tokens per second with MTP-1, and approximately 60 tokens per second with DSpark. A five-run mean recorded 60.31 tokens per second, a 1.51x gain over MTP-1.

Under heavy server load with a strict 120 token-per-second latency requirement, V4-Flash achieved a 661% throughput gain. At an 80 token-per-second requirement, the gain was 51%. The difference reflects how DSpark dynamically shifts between optimizing for single-request speed and multi-request throughput based on observed GPU conditions.

Companion Toolkit and Licensing

DeepSeek also released DeepSpec, a full-stack codebase for training and evaluating speculative decoding draft models. DeepSpec currently supports three draft model types, DSpark, DFlash, and Eagle3, and two target model families, Qwen3 and Gemma. Both releases are available under the permissive MIT license.

Before DSpark, speculative decoding tooling was scattered across individual research teams. DeepSpec consolidates data preparation, draft model implementations, training code, and evaluation scripts into a single reproducible toolchain.

Business Implications

For software companies running AI workloads at scale, an 85% speed increase without hardware procurement changes means lower per-token costs and better latency for end users. Teams evaluating on-premise or private cloud AI deployments can now access production-grade inference optimization without proprietary dependencies.

The “DSpark proves that software innovation can compensate for hardware limitations,” the DeepSeek research team stated. Liang Wenfeng said in a prior interview: “Money has never been the problem for us; bans on shipments of advanced chips are the problem.” DSpark reflects that pressure, delivering more output from existing silicon.

Read more: DeepSeek open sources DSpark, a new framework to speed up LLM inference by up to 85%