Morning Edition · Sunday, June 28, 2026

DeepSeek Open-Sources DSpark, Claiming Up to 4x Throughput on V4 via Speculative Decoding

The semi-parallel decoding method ships as an add-on module to the existing V4 Flash and Pro checkpoints, and DeepSeek also released code for training draft models on other architectures.

Save

DeepSeek Open-Sources DSpark, Claiming Up to 4x Throughput on V4 via Speculative Decoding

DeepSeek released DSpark, a speculative-decoding method that the lab says raises throughput on its V4 Flash and V4 Pro models by 51 to 400 percent depending on concurrency and use case. The enhanced checkpoints, which pair the original base model with an attached DSpark module, are already published.

Measured against the model's existing multi-token-prediction baseline, DeepSeek reports per-user generation gains of 60 to 85 percent on Flash and 57 to 78 percent on Pro, according to coverage of the release. The shipped configuration, DSpark-5, uses a five-token draft block with a Markov head, and the lab also open-sourced DeepSpec, the codebase for training and evaluating draft models, with reported transfer to other model families such as Gemma and Qwen.

The wide range, 51 to 400 percent, is the part to read carefully. The top of that band reflects favorable batch and concurrency settings rather than a guaranteed result, and the figures are the lab's own. Speculative decoding does not change output quality when verification is exact, so the claim is about cost and latency, not capability. The practical significance is that a frontier-class open-weight model now serves more tokens per accelerator at no cost to quality, which directly lowers the cost of running it outside the major clouds.

Veracity: Corroborated

77/100

If true, who benefits: DeepSeek and operators serving open-weight models outside the major clouds, who recover throughput that chip-access limits would otherwise deny them.
The nuance: The 51 to 400 percent range is the lab's own figure, and the top of the band reflects favorable concurrency settings rather than a result independent deployments have reproduced.

An open-source-intelligence read of how likely this story is true with its real nuance, not a judgment of any outlet. It assesses the claim, weighing independent and adversarial reporting. How we label confidence.

What this means

Inference economics, not just model quality, are now an area of competition for open-weight labs, and getting more throughput from the same hardware partly offsets the access restrictions that limit who can buy frontier accelerators. For anyone serving DeepSeek-V4, the release is an immediate way to cut costs rather than a future research direction.

What to watch

Independent throughput reproductions on standard hardware, which will show whether real deployments see the high end of the claimed range or the conservative 51 percent floor.
How fast the open-sourced DeepSpec draft-model recipe is applied to other open-weight families, a signal that the gain generalizes rather than being specific to DeepSeek's architecture.

Observations to monitor, not financial advice.

2 sources

Synthesized from: Polylog editors · MarkTechPost

Part of a tracked trend

Software-Level Inference Efficiency

Decoding and serving optimizations keep cutting the cost of running a given model without retraining it, steadily lowering the compute barrier to deploying frontier-class open-weight models and partly neutralizing hardware access limits.