DeepSeek launches DSpark: up to 85% faster V4 inference

DeepSeek has unveiled DSpark, a speculative decoding framework aimed at speeding up inference for its DeepSeek-V4 model series. According to the official Hugging Face model card, DeepSeek-V4-Pro-DSpark is not a new model: it uses the exact same V4-Pro checkpoint with an added speculative decoding module. DeepSeek says the technique speeds up inference by up to 85%.

Speculative decoding uses a small draft model to guess several tokens ahead, then has the large target model verify those tokens in parallel. By cutting the number of sequential decoding steps, token generation gets faster while output quality is preserved — an important optimization as GPU costs and infrastructure strain keep rising.

Key points

• DSpark = a speculative decoding module attached to the V4-Pro checkpoint, leaving the base model weights unchanged.
• DeepSeek claims inference speedups of up to 85%, easing the inference workload.
• DSpark is one algorithm inside DeepSpec — a full-stack open-source codebase for training and evaluating draft models: data preparation, draft model implementations, training and evaluation.
• The repo ships a config for Qwen3-4B and evaluates on benchmarks such as gsm8k, math500, aime25, humaneval, mbpp, livecodebench, mt-bench, alpaca and arena-hard-v2.
• V4 context: V4-Pro has 1.6T parameters (49B activated), V4-Flash 284B (13B activated), both supporting a 1M-token context with a hybrid attention architecture (CSA + HCA).

For developers, the notable part is that DeepSeek open-sourced the draft-model training pipeline rather than just publishing results — meaning other teams can train draft models for their own models using the same recipe.

FAQ

Is DSpark a new model? No. The model card states it is the same V4-Pro checkpoint with a speculative decoding module added.

Why is it faster without losing quality? A small draft model proposes several tokens that the target model verifies in parallel; mismatched tokens are rejected, so the final output matches standard decoding.

Is it open source? Yes. DSpark lives inside DeepSpec — a complete codebase (data prep, training, evaluation) published on GitHub with sample draft-model configs.