speculative-decoding — AI Digest

28 авг Tencent open-sources Hy4 preview: 770B / 49B-active MoE with Gated DSA + iHC, 1M context, Apache 2.0 Tencent models-llm
26 июн JetSpec: Breaking the Scaling Ceiling of Speculative Decoding with Parallel Tree Drafting Hao AI Lab, UC San Diego research
29 июн DeepSeek Open-Sources DSpark: 57–85% Inference Speedup for V4 in Production DeepSeek tools
29 июн JetSpec: Parallel Tree Drafting Achieves 9.64× Speculative Decoding Speedup Hao AI Lab, UCSD research
28 авг vLLM v0.28.0 ships Kimi-K3 perf push, full DeepSeek-V4 sparse-MLA, Model Runner V2 maturity, and tiered KV-cache offloading vLLM Project tools
6 мая SGLang v0.5.11: Speculative Decoding V2 as Default and Eight New Model Architectures tools
16 мая Orthrus: 7.8x Inference Speedup for Qwen3 via Autoregressive-Diffusion KV Sharing research
18 мая vLLM v0.21.0: Blackwell MLA Backend, HMA KV Offload, Spec Decode for Reasoning Models vLLM Project tools
2 июл BlockPilot: Instance-Adaptive Block Size for Diffusion-Based Speculative Decoding research
6 мая Ollama v0.23.1: Gemma 4 MTP Speculative Decoding Delivers 2× Speed on Apple Silicon tools
17 июн llama.cpp June 16 Builds: Eagle3 Speculative Decoding, Vulkan UMA Memory, NVFP4 Fixes tools
29 июн llama.cpp Builds b9830–b9837: DFlash v2, MiniCPM5 Parser, --reasoning-preserve Flag ggml-org tools
14 июл llama.cpp Adds Tencent Hunyuan 3, Minimax2 Eagle3 Speculative Decoding, and SYCL Fused MoE tools
1 сен llama.cpp brings MTP speculative decoding to recurrent Qwen models ggml tools
17 окт Tencent opens up DSpine parallel speculative-decoding drafter Tencent research