DeepSeek open-sources DSpark speculative decoding and the DeepSpec codebase

Category: AI
Status: developing
Sources: DeepSpec repo, DSpark paper, model card, discussion
Summary: On 2026-06-26 DeepSeek published DeepSpec, an MIT-licensed full-stack codebase for training and evaluating speculative-decoding draft models, alongside DSpark, a draft module attached to DeepSeek-V4 checkpoints. Speculative decoding is lossless, so output is identical to standard decoding; the model card states DeepSeek-V4-Pro-DSpark is not a new model but the same checkpoint with a speculative-decoding module attached. The repository also implements the DFlash and Eagle3 draft models, evaluates over gsm8k, math500, aime25, humaneval, mbpp, and livecodebench, and trains drafts for non-DeepSeek targets including Qwen3 and Gemma. The Hacker News submission cites 60 to 85 percent faster generation; these are the project's own figures and are not independently reproduced.
Why it matters: A lossless, open-source inference speedup that transfers across model families lowers serving cost for self-hosted LLM deployments.
Follow-up: Watch for independent throughput reproduction, integration into serving stacks such as vLLM and SGLang, and per-target acceptance-length numbers.