vLLM v0.25.0 released with Model Runner V2 as default and PagedAttention removed
vLLM v0.25.0 is now available, featuring 558 commits from 232 contributors. Model Runner V2 becomes the default execution path for all dense models, with new support for EVS, realtime embeddings, prefix caching for Mamba hybrid models, multimodal-prefix bidirectional attention, and dynamic speculative decoding. The legacy PagedAttention implementation has been removed.
Entities
Related
Hugging Face releases transformers v5.13.1 patch to enable vLLM compatibility
Hugging Face released transformers v5.13.1, a patch focused on enabling compatibility with the latest vLLM release. The update includes defensive handling of legacy layer types, fixes for custom code with new linear layer names, and a fix for _LazyAutoMapping.register with string keys.
Ollama releases v0.32.0-rc0 with Qwen3.5 support and agent UI
Ollama released v0.32.0-rc0, adding a Qwen3.5 parser and renderer, a warning for old agent models, and a new agent UI. This update enables users to work with Qwen3.5 models and improves the agent experience.
Software engineer publishes final part of LLM-from-scratch series covering inference and decoding
A software engineer published the fourth and final part of a blog series explaining LLMs from the ground up, focusing on token-by-token generation, KV cache, and decoding strategies (temperature, top-k, top-p). The series aims to help other software engineers understand the internals of LLMs.
AgentMaker: a new Python framework for building LLM agents and multi-agent systems
AgentMaker is a general-purpose Python framework for building LLM agents and multi-agent systems, featuring tools, memory, RAG, context engineering, guardrails, human-in-the-loop, and observability. It is released under MIT license on GitHub and PyPI.
LLM hardware recipe database with filters and community usage tracking
A community-driven database that lists which LLM models run on which hardware, with performance details. Users can filter by hardware, submit new recipes, and mark which recipes they actively use to show popularity. It solves the problem of finding compatible model-hardware combinations for LLM deployment.