DeepSeek shares DeepSpec tools for faster model output
DeepSpec is a full codebase for training and testing used in . lets a smaller model suggest upcoming text first, while a larger it, which can make generation faster. The collection includes data preparation tools, draft model code, training code, and evaluation scripts.
The released checkpoints are the ones used for Table 1 in the paper. Each checkpoint was trained on open-perfectblend data made by its matching in non-, and each one comes from the training setup in the repository config. The include Qwen3 4B, 8B, 14B, and Gemma 4 12B variants, with Eagle3, DFlash, and DSpark algorithms.
Anyone citing or comparing these results needs to match the repository’s training settings, or the comparison may not be meaningful.
Key points
- DeepSpec provides code for training and evaluating for .
- It includes data tools, model s, training code, and evaluation scripts.
- Released checkpoints cover Qwen3 and Gemma 4 .
- Eagle3, DFlash, and DSpark algorithm checkpoints are included.
- Comparisons should use the same training settings as the repository.