DeepSeek

Type: Organization (model lab operating its own serving stack)

DeepSeek matters on this rung for a reason unrelated to its models: it is the only lab here that has published the serving system it actually runs in production. DSpark is deployed in the DeepSeek-V4 serving stack and measured on live traffic, not on a benchmark harness.

The system couples parallel semi-autoregressive drafting with confidence-scheduled, load-aware verification — the scheduling half is the novel part. Prior parallel drafters propose long token sequences in one forward pass but suffer rapid acceptance decay from missing inter-token dependencies, and verifying those extended blocks indiscriminately wastes batch capacity on tokens likely to be rejected. DSpark makes verification depth a function of confidence and current load.

Key contributions

• 60–85% faster per-user generation on V4-Flash and 57–78% on V4-Pro against the prior production baseline (MTP-1), at matched throughput (DSpark)
• Retains the exact distribution-preserving guarantee — the speedup is not paid for in output quality (DSpark)
• Makes verification load-aware, tying draft acceptance to live batch pressure rather than a fixed depth (DSpark)

Evidence caveat

First-party self-measurement of the authors' own production system, published as a preprint. There is no independent reproduction, and the baseline (MTP-1) is also DeepSeek's. The close-read grades it moderate at best — the numbers describe a real deployment, which is rare and valuable, but they are unaudited.

Mentioned in

• Speculative Decoding — the only production deployment in this rung
• Frontier — cited under the independent-reproduction gap