Context Engineering Model serving Inference Optimization KV Cache GPU Infrastructure vLLM Cost Optimization Benchmarks Kimi MiniMax DeepSeek V4 Flash

DSpark

Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation

Visit site

DSpark is DeepSeek-AI's speculative decoding framework/drafter approach for accelerating LLM inference. It combines semi-autoregressive parallel draft generation with confidence-scheduled, load-aware verification, and DeepSeek released DSpark checkpoints and implementations through the DeepSpec repository.

Recent stories

4 linked stories

releaseSECONDARY2026-10-22

vLLM adds NVFP4 serving for MiniMax M3.1 Flash at 76 concurrent sequences

vLLM added NVFP4 weights, NVFP4 KV caching and DSpark support for MiniMax M3.1 Flash. It reports 76 concurrent sequences at 256K context on four B300 GPUs, versus 28 for M3.

releaseSECONDARY2026-10-17

vLLM reports 842 tokens per second for Kimi decoding with NVIDIA DSpark

vLLM reports 842 tokens per second at batch one on GB300 NVL72 with DSpark, plus up to 1.8× better P90 interactivity. Engineers dispute describing the implementation as a megakernel.

workflowPRIMARY2026-08-12

Speculative decoding tests report acceptance drop from 0.71 to 0.18 after ~32K context

A practitioner report found speculative-decoding acceptance fell from 0.71 to 0.18 beyond about 32K context. Separate DSpark and mlx-dspark tests reported speedups on RTX and Apple Silicon setups.

releaseSECONDARY2026-08-03

DeepSeek V4 Flash adds Baseten and Together AI serving with 1M-token context

Baseten and Together AI added DeepSeek V4 Flash with a 1M-token context window, reasoning-effort controls, and DSpark decoding. ValsAI ranked it the cheapest model above 60 on its index, and Nous promoted a short 90% discount.