DeepSeek DSpark: How the Open-Source Speculative Decoding Framework Speeds Up V4 by Up to 85%

DSpark introduces a novel element called confidence-scheduled speculative decoding. The system dynamically decides how many tokens to speculate based on confidence levels, which reduces wasted verification computation . It replaces DeepSeek-V4's previous MTP-1 (Multi-Token Prediction) scheme in production .

Production Performance Gains

DSpark is already deployed on DeepSeek-V4's production systems handling real user traffic on the DeepSeek-V4-Flash preview and DeepSeek-V4-Pro preview services . At the same total system throughput, DSpark delivers the following per-user generation speed improvements compared to the prior MTP-1 baseline:

These results come from live user traffic, not synthetic benchmarks . Under tight latency constraints, DSpark avoids the throughput cliff that earlier schemes suffered, pushing out the system's Pareto frontier . In one test targeting 120 tokens/second/user for V4-Flash, MTP-1 was near its capacity limit while DSpark delivered a nominal throughput advantage of 661% .

Model-Agnostic Compatibility

DSpark is designed to be model-agnostic. The paper demonstrates its effectiveness across non-DeepSeek architectures: on Qwen3-4B, Qwen3-8B, and Qwen3-14B, DSpark improved the macro-average accepted length by 30.9%, 26.7%, and 30.0% respectively over the Eagle3 baseline . Against the parallel draft model DFlash, gains were 16.3%, 18.4%, and 18.3% on the same Qwen3 sizes . DSpark also maintained its lead on Gemma4-12B . Notably, a 2-layer DSpark configuration outperformed a 5-layer DFlash configuration .

Scaling draft length from 4 to 16 tokens added only 0.2–1.3% per-round latency, while accepted length improved by up to 30% .

The DeepSpec Codebase

Alongside DSpark, DeepSeek open-sourced DeepSpec, a full-stack speculative decoding training and evaluation framework. It includes implementations of Eagle3, DFlash, and DSpark, and enables developers and researchers to:

• Reproduce results reported in the DSpark paper
• Train custom draft models
• Evaluate speculative decoding pipelines
• Extend the approach to their own models

The paper, code, and model weights are hosted under the deepseek-ai/DeepSpec repository on GitHub and on Hugging Face .

Broader Context: V4 Official Launch & Peak/Off-Peak Pricing

On June 29, 2026, DeepSeek announced that the official DeepSeek V4 release is scheduled for mid-July 2026 . Alongside it, DeepSeek will introduce a peak-and-off-peak (time-of-day) API pricing structure :

• Peak hours: 9:00–12:00 and 14:00–18:00 Beijing time daily .
• Pricing: API costs double during peak hours relative to off-peak (standard) rates .
• Example (V4-Pro, per million tokens):
  • Input (cache hit): 0.025 RMB off-peak → 0.05 RMB peak
  • Input (cache miss): 3.00 RMB off-peak → 6.00 RMB peak
  • Output: 6.00 RMB off-peak → 12.00 RMB peak

For V4-Flash, the corresponding peak pricing doubles from 0.02 RMB to 0.04 RMB (cache hit), 1 RMB to 2 RMB (cache miss), and 2 RMB to 4 RMB (output) per million tokens . DeepSeek stated the change is meant to "more rationally allocate resources and improve service stability" . Users will receive email notifications 24 hours before billing changes take effect . This pricing shift, combined with the DSpark speedups, signals DeepSeek's push to balance commercial monetization (post its approximately 50 billion RMB funding round) with continued aggressive open-source releases .