← Back to feed

DeepSeek on June 27 released DSpark, an open-source speculative decoding framework that accelerates inference on its V4 models by 60 to 85% without retraining the underlying weights. The framework is live in production on DeepSeek’s API, available as MIT-licensed code on HuggingFace and GitHub, and works with third-party model families including Qwen3 and Gemma.

The headline numbers from DeepSeek’s internal production data: V4-Flash sees 60–85% faster per-user generation versus the prior MTP-1 baseline; V4-Pro sees 57–78%. Aggregate throughput at matched system capacity improved 51–52% for both models. Offline testing showed an accepted token length improvement of 26.7–30.9% over Eagle3 and 16.3–18.4% over DFlash.

How It Works

Standard speculative decoding drafts multiple candidate tokens using a small, fast model, then verifies them against the full model in a single forward pass. When the draft is accurate, you get multiple tokens for approximately the cost of one verification step. Quality is lossless: the output distribution of the target model is preserved regardless of speculative failures.

DSpark’s two contributions are about where standard speculative decoding wastes compute.

The first is architecture. Rather than a separate draft model, DSpark grafts the speculative head directly onto the target model. The drafter shares embeddings and the LM head with the full model and trains only three small auxiliary modules: a parallel backbone, a Markov serial correction head, and a confidence head. This requires 27% of the single-token inference FLOPs and 10% of the KV cache used by DeepSeek V3.2 at a 1M-token context — meaning the draft cost is genuinely cheap.

The second is confidence-scheduled verification. Classic speculative decoding sends all draft tokens to the target model regardless of whether they are likely to be accepted, burning verification budget on certain rejections. DSpark’s confidence head scores each draft token’s probability of acceptance. A hardware-aware prefix scheduler uses those scores to dynamically trim verification length: longer verification prefixes when GPUs are idle, shorter ones under load. This eliminates wasted verification cycles without touching the lossless guarantee.

The deployed production configuration is DSpark-5 — five-token draft blocks with the Markov correction head. That is the configuration behind the published production numbers.

What Ships

Two HuggingFace checkpoints: DeepSeek-V4-Flash (284B total parameters, 13B active) and DeepSeek-V4-Pro (1.6T total parameters, 49B active), both with 1M-token context windows. The DeepSpec codebase covers training and evaluation for DSpark modules and supports V4-Flash, V4-Pro, Qwen3 families, and Gemma families under the same MIT license. Target model weights are not modified.

An independent developer, Rafael Caricio, documented single-stream V4-Flash benchmarks: 26.3 tokens per second without speculative decoding, 39.9 tokens per second with MTP-1, and approximately 60 tokens per second with DSpark. That is roughly 1.5x over MTP-1 and 2.3x over unaccelerated baseline on a single-stream workload.

Why It Matters

DeepSeek V4-Pro is already the cheapest per-output-token option among frontier-tier models, at $0.87 per million output tokens after its permanent 75% price cut. DSpark does not change that price. What it changes is the compute cost to serve those tokens — and therefore DeepSeek’s margin per request and the maximum concurrency the same hardware can handle.

For anyone self-hosting V4 weights, the math is direct: the same GPU cluster now handles 1.5x to 4x the request volume depending on concurrency level. The throughput ceiling that made 1T+ parameter models impractical on smaller clusters moves substantially.

For third-party providers serving Qwen3 or Gemma, the DSpark codebase is the fastest path to production-grade speculative decoding on those architectures today.

The main caveat: all speedup numbers are self-reported by DeepSeek. Production outcomes vary with workload, serving configuration, and traffic patterns. The accepted-token-count advantage over Eagle3 is the most verifiable number, and it has already been roughly replicated by independent developers on single-stream workloads. Whether the 85% peak figure survives a mixed-workload production deployment is the open question.