Researchers at the University of California, San Diego have introduced DFlash, an open-source speculative decoding framework engineered to accelerate large language model inference on NVIDIA Blackwell and Hopper GPU architectures. As generative AI workloads transition toward coordinated multiagent systems, the sequential token generation of autoregressive decoding has become a critical bottleneck for latency-sensitive serving. Unveiled in February 2026, DFlash addresses this constraint through a novel block-diffusion drafting mechanism that parallelizes token prediction without compromising output fidelity. Traditional speculative decoding relies on lightweight autoregressive models to draft tokens sequentially before a larger target model verifies them in parallel. DFlash replaces this sequential drafter with a block-diffusion model capable of predicting an entire block of masked future tokens in a single forward pass. The target model subsequently verifies these candidates simultaneously. This architectural shift transforms sequential drafting into block-parallel GPU operations, allowing systems to bypass memory-bound decode phases and fully utilize the high-bandwidth compute capabilities of modern data center accelerators. Independent benchmarks demonstrate substantial performance gains across diverse workloads. When deployed on an eight-NVIDIA DGX B300 system utilizing TensorRT-LLM, DFlash increases inference throughput for the gpt-oss-120b model by up to fifteen times at interactive latency targets. The framework also nearly doubles per-user interactivity for Llama 3.1 8B at equivalent concurrency levels, consistently outperforming established methods such as EAGLE-3 across coding, retrieval-augmented generation, reasoning, and multilingual datasets. On single NVIDIA Blackwell Ultra GPU deployments, DFlash achieves up to 5.8x throughput improvements on Gemma 4 31B and 5.1x on Qwen3 8B within production benchmarks. Rapid integration into developer workflows is already underway. The research team has published twenty DFlash model checkpoints on Hugging Face, covering major architectures including Qwen, Llama, Gemma, Kimi, and gpt-oss. NVIDIA and the open-source community have aligned framework support across vLLM, SGLang, and TensorRT-LLM. Integration requires minimal configuration adjustments and eliminates the need for application refactoring. Developers can deploy DFlash via the open-source Speculators library, which routes draft model proposals directly into the target model hidden states within the standard inference pipeline. The framework is now optimized for both NVIDIA Blackwell and Hopper architectures, leveraging the unified compute domain and high-speed chip-to-chip interconnects inherent to next-generation accelerators. By decoupling drafting from sequential token prediction, DFlash establishes a new efficiency baseline for agentic AI and real-time conversational systems, enabling enterprises to scale concurrent user sessions while maintaining strict latency service-level agreements.