DeepSeek releases DeepSpec: open-source full-stack for speculative decoding

On 2026-06-26, DeepSeek created a public GitHub repository at deepseek-ai/DeepSpec — a full-stack MIT-licensed codebase for training and evaluating speculative-decoding draft models (README, 2026-06-28). As of 2026-06-28 the repo sits at 1,546 stars, 128 forks, with zero release tags (GitHub REST API, 2026-06-28). Every speedup number in the article is the authors’ own measurement — no independent reproduction exists.

What it is

DeepSpec ships three draft-model implementations in one repo (README, 2026-06-28):

• DSpark — DeepSeek’s own new draft model. Paper ships in the repo as DSpark_paper.pdf; no head-to-head number against DFlash or Eagle3 in the README.
• DFlash — block-diffusion draft model. Per the paper: “DFlash achieves over 6x lossless acceleration across a range of models and tasks, delivering up to 2.5x higher speedup than EAGLE-3” (arXiv:2602.06036, 2026-06-28). Accepted at ICML 2026 (camera-ready).
• Eagle3 — prior state of the art. Per its paper: “speedup ratio up to 6.5x, with about 1.4x improvement over EAGLE-2” (arXiv:2503.01840, 2026-06-28).

The pipeline runs in three stages: data preparation (downloads prompts, regenerates target answers, builds target cache — ~38 TB for the default Qwen3-4B), training (one worker per visible GPU, default 8 GPUs), and evaluation (nine benchmarks: gsm8k, math500, aime25, humaneval, mbpp, livecodebench, mt-bench, alpaca, arena-hard-v2) (README, 2026-06-28).

Why it matters

A “training day” for speculative-decoding stacks. Before DeepSpec, building a custom draft model meant reading the Eagle3 paper, finding unofficial implementations, writing your own training loop, and reconciling target-cache formats. DeepSpec consolidates the four steps into one repo.

DFlash is an ICML 2026 paper. The academic marker is on the record. Camera-ready acceptance is not a benchmark reproduction, but it is the second-strongest signal in the speculative-decoding literature for the DFlash technique.

The architectural difference. DFlash drafts multiple tokens in a single block-diffusion forward pass; Eagle3 drafts token-by-token at the feature level. The “up to 2.5×” is block-diffusion drafting vs autoregressive-feature drafting in the same setup.

The hardware floor is real. 8 GPUs as default, 38 TB target-cache storage for Qwen3-4B. The README states both explicitly.

Supported targets are Qwen3 and Gemma — not DeepSeek’s own models. The article does not imply DeepSeek uses DeepSpec in production for DeepSeek-V3 or DeepSeek-R1.

Practical implications

• Evaluating speculative-decoding stacks: Compare DSpark vs DFlash vs Eagle3 on the same target, eval set, and hardware. DeepSpec provides the harness.
• Training a custom draft model: 8-GPU / 38 TB budget is the planning number for Qwen3-4B.
• Studying speculative decoding: Three algorithms, three papers, one repo — the previous fragmentation is resolved.
• Wiring into a serving stack: Only Qwen3 and Gemma targets are supported. Not DeepSeek-V3/R1 or Llama-3 at this commit.

Risks and caveats

1. 6× and 2.5× are the DeepSeek authors’ own measurements. No independent reproduction exists as of 2026-06-28. Attribute them as authors’ claims.
2. DSpark has no headline benchmark in the README. Paper in the repo, no published head-to-head against DFlash or Eagle3.
3. Supported targets: Qwen3 and Gemma only. Do not imply DeepSeek uses DeepSpec for its own models.
4. Hardware floor: 8 GPUs, 38 TB storage. Real and named in the README.
5. The repo is one day old. 10 commits, no releases, no v0.1.0. Breaking changes are a real possibility.
6. “Up to 2.5×” is doing work. Expect a per-workload spread, not a flat win.
7. DSpark’s paper contents are not separately verified.
8. 1,546 stars is a launch spike. Star counts at this band move by the thousand per week.

What to watch

1. A first tagged release (currently zero releases)
2. Independent reproductions of the 6× and 2.5× numbers
3. Support for additional target models (DeepSeek-V3, DeepSeek-R1)
4. DSpark’s standalone performance (first benchmark)
5. Adoption signals beyond the launch spike (blog posts, engineering write-ups)
6. A camera-ready v3 of the DFlash paper

Verdict

DeepSpec is a real, working, one-day-old open-source speculative-decoding training stack: three draft-model implementations (DSpark, DFlash, Eagle3), 8-GPU training scripts, 38 TB target-cache storage, and evaluation across nine benchmarks. DFlash’s ICML 2026 acceptance is the academic marker; the 6× and 2.5× claims are the authors’ own. The first tagged release, first independent reproduction, and first DSpark head-to-head are the next-cycle signals (README, 2026-06-28; GitHub REST API, 2026-06-28).

Source table