毎日AIニュース 0628 毎日AIニュース 0628

DeepSeek and Peking University have released DSpark, an open-source framework for speculative decoding aimed at accelerating large language model inference, with the teams reporting that it raised generation speed by as much as 85 percent in DeepSeek-V4 production deployments while holding throughput steady. The release matters because inference cost and latency remain the central bottleneck for deploying capable models at scale, and a freely available optimization framework could lower the barrier for both cloud providers and on-device or local-LLM operators.

Speculative decoding is the technique at the core of DSpark. In a conventional autoregressive setup, a model generates one token at a time, and each step requires a full forward pass, which makes latency scale linearly with output length. Speculative decoding instead pairs a small, fast "draft" model with the larger target model. The draft proposes several tokens ahead, and the larger model verifies them in a single batched pass, accepting the ones that match what it would have produced and discarding the rest. Because verification is cheaper than sequential generation, this can yield large speedups without changing the final output distribution, which is why the approach has gained traction across the industry. DSpark appears to package these ideas into a production-ready toolkit rather than introduce an entirely new method.

The reported figures—up to 85 percent faster generation while maintaining throughput—are notable because the two metrics often trade off against each other. Throughput, typically measured in tokens per second across many concurrent requests, can suffer when systems chase lower single-request latency. Maintaining both suggests careful engineering around batching, draft acceptance rates, and memory management, though the gains will likely depend heavily on workload, hardware, and the gap between draft and target model sizes. As with most vendor benchmarks, independent reproduction will be the real test of how broadly the 85 percent figure holds.

The collaboration between DeepSeek and Peking University reflects a broader pattern of Chinese labs and universities pushing capable models and tooling into the open. DeepSeek built its reputation with the V2 and V3 series and the R1 reasoning model, which drew attention for strong performance at comparatively low training cost. Open-sourcing the inference framework behind DeepSeek-V4, rather than only the weights, extends that posture and could pressure competitors to match the transparency. It also fits the local-LLM theme of this report, since faster inference frameworks are exactly what hobbyists and small teams need to run large models on constrained hardware.

DSpark enters a crowded field of inference optimizations. vLLM popularized PagedAttention for efficient memory use, while TensorRT-LLM, Hugging Face's Text Generation Inference, and SGLang all target high-throughput serving. Speculative methods specifically include Medusa, EAGLE, and Lookahead decoding, alongside the original draft-model approach. Where DSpark fits among these will depend on its integration, supported hardware, and whether it works cleanly with mixture-of-experts architectures, which DeepSeek favors and which complicate batching and routing. If the framework is model-agnostic enough to accelerate other open models, adoption could spread quickly; if it is tuned mainly for DeepSeek's own stack, its impact may be narrower.

For readers tracking the practical side, a few caveats are worth keeping in mind. Speculative decoding shines when the draft model agrees with the target frequently, so accuracy gains tend to be largest on predictable text and smaller on highly uncertain generation. The technique also adds engineering complexity and memory overhead from running two models, which can be a constraint on local setups. Quantization, KV-cache optimization, and continuous batching remain complementary levers that teams typically combine for the best results.

The wider takeaway is that the competitive frontier in 2026 is shifting from raw model quality toward efficiency and serving economics. By releasing DSpark openly, DeepSeek and Peking University appear to be betting that ecosystem adoption is more valuable than keeping the optimization proprietary. Whether DeepSeek-V4 itself becomes broadly available and how the framework performs outside the lab's own infrastructure are the open questions that will determine how meaningful this release proves to be.