vllm.aiCrawled Oct 29, 2026
Complete archive of all vLLM blog posts in chronological order.
Technical articles, release announcements, model guides, and community updates from the vLLM project.
Day-0 vLLM support for Mistral Large 4 with sparse MLA, cross-layer KV sharing, DSpark speculative decoding, DCP/PCP long context and NIXL P/D on NVIDIA Blackwe
vLLM serves GLM-5.3 on 16 H200s with PCP and DCP prefill and Hybrid HiSparse decode, reaching 1M-token contexts and 1.7–3.2× the per-user speed of data-parallel
The page lists all vLLM releases and provides installation commands and release notes for each version.
A performance comparison of vLLM and DeepSpeed-FastGen, explaining when Dynamic SplitFuse helps, where vLLM is faster, and how memory allocation, output length,
The vLLM team invites inquiries for collaborations, talent recruitment, social media partnerships, and website feedback.
A bi-weekly livestream with vLLM committers: project updates, guest talks, and open Q&A. Includes recordings and slides from past sessions.
vLLM is an inference and serving engine designed for high-throughput, memory-efficient deployment of large language models.
FuriosaAI Inc. secures **$620 million** in funding to scale production of its energy-efficient AI inference chips, targeting data center power constraints. The startup’s **Tensor Contraction Processor (TCP)**, based on its **RNGD architecture**, uses a **three-dimensional matrix** for tensor computations, reducing power consumption by **less than half** compared to Nvidia’s GPUs while maintaining compatibility with frameworks like PyTorch. Major South Korean firms—**LG Group, Upstage, and Samsung**—already deploy its chips for AI models, live search summaries, and neural processing services. The funding will accelerate production, aiming for **20,000 units in 2026** and **40,000 in 2027**, while advancing a **third-gen processor** co-developed with Broadcom.
Oct 2026SiliconANGLENews article
