# deepseek-ai/DeepSpec

> Full-stack pipeline for training and evaluating draft models for speculative decoding: data preparation, training against a target-model cache, and acceptance measurement on benchmarks. Supports DSpark, DFlash and Eagle3.

- Magnitude: 1.1 out of 10 — Steady
- Stars: 7,260 total · +6 stars measured 2026-10-29, 00:11–17:04 UTC, ≈ 8 by evening
- Star trust: star growth looks organic
- Category: Language models · Language: Python · License: MIT · Created: 2026-06-26 · Last push: 2026-07-09
- GitHub: https://github.com/deepseek-ai/DeepSpec · Page: https://gitnova.dev/en/r/deepseek-ai/DeepSpec

## Useful for

- Train a draft model to speed up Qwen3-4B inference on a single 8-GPU node
- Measure speculative-decoding acceptance on gsm8k and humaneval with a released checkpoint
- Build a target cache from open-perfectblend to train a new draft model

## Why it’s here

- Star-counter measurements on 2026-10-29 (UTC), 00:11–17:04: 7254 → 7260 stars (+6). This is the change over that interval.
- Estimated end-of-day forecast: about +8 stars, using observed gains and the previous day.

## Star trust

Star growth looks organic.

Star-trust labels are heuristics based on the repository’s behavior, not a check of every stargazer.

## Numbers

- Forks: 693
- Issues and pull requests: 88
- Watchers: 34
- Average over the last week: 4 per day
- Usual pace: 4 per day
- Stars in the last hour (measured): 0

## Stars per day, last 30 days (oldest → newest, today is partial)

2026-09-30 … 2026-10-29: 6, 2, 1, 3, 3, 6, 2, 2, 4, 4, 2, 6, 8, 1, 2, 5, 1, 2, 4, 2, 11, 4, 2, 4, 1, 3, 6, 4, 1, 6

## Magnitude by day, last 14 days

10-16 1.1, 10-17 0.8, 10-18 0.8, 10-19 1.1, 10-20 1.1, 10-21 1.6, 10-22 1.2, 10-23 0.9, 10-24 1.2, 10-25 0.8, 10-26 1.2, 10-27 1.5, 10-28 1.2, 10-29 1.1

## Similar by description

1. **deepseek-ai/FlashMLA** — 1.4 · Steady · Language models · C++ · +4 stars measured 2026-10-29, 00:09–17:02 UTC, ≈ 6 by evening
   DeepSeek's library of optimized attention kernels (sparse and dense) for running DeepSeek-V4.1 inference on NVIDIA GPUs and Huawei Ascend NPUs, with FP8/FP4 KV cache support.
   Full card: https://gitnova.dev/en/r/deepseek-ai/FlashMLA.md

---

Magnitude (0–10) measures how fast and how unusually interest in a repository is growing right now. It is not a quality score.
Days are UTC. “So far today” is a fact; “expected by the end of the day” is a forecast.
Summaries and use cases are written by an LLM (DeepSeek V4.1 Flash) from the README and may be inaccurate: verify specific claims (benchmarks, speed, hardware) in the repository itself.
Data as of 2026-10-29 17:07 UTC, updated every 30 minutes.