DSpark
#77 in Open-Source LLMsDeepSeek · since 27. Juni 2026 · 8× · last seen Jul 05, 2026
15
Momentum
DSpark is an open-source speculative decoding framework developed by DeepSeek together with Peking University that speeds up large language model inference without changing the underlying model's output. It is not a standalone model but an additional draft module attached to existing DeepSeek-V4-Pro and DeepSeek-V4-Flash checkpoints. The code, along with the DeepSpec training framework, was released under the MIT license on GitHub and Hugging Face on June 27, 2026. DeepSeek reports self-measured, not yet independently verified, gains of 60–85% faster per-user token generation versus the prior MTP-1 baseline.
Momentum trend
19.05.17.08.
Features
| Key-Benchmark (%) | 60–85% schnellere Token-Generierung pro Nutzer ggü. MTP-1-Baseline (V4-Flash); 57–78% (V4-Pro); Akzeptanzlänge +26,7–30,9% ggü. Eagle3, +16,3–18,4% ggü. DFlash |
| Kontextfenster (Token) | 1.000.000 Token (geerbt von DeepSeek-V4-Pro/Flash, auf denen DSpark aufsetzt) |
| Lizenz | MIT-Lizenz |
| Multimodalität | Nicht multimodal – reines Text-Inferenz-Beschleunigungsframework (Speculative Decoding) für Sprachmodelle |
| Plattform | GitHub (deepseek-ai/DeepSpec) und Hugging Face (DeepSeek-V4-Pro-DSpark, DeepSeek-V4-Flash-DSpark); Serving-Integration u.a. via vLLM, SGLang, vLLM-Ascend |
| Release-Datum | 27. Juni 2026 |