VDF.AI

Open-source RAG frameworks checked against their GitHub repositories and documentation in October 2026: LlamaIndex, LangChain with LangGraph, Haystack, RAGFlow, DSPy, txtai, Kotaemon and LightRAG, compared on ingestion, retrieval, evaluation, agentic RAG and deployment, with the projects that are archived or in maintenance mode flagged.

Which open-weight model to run on an 8, 12, 16, 24 or 32 GB graphics card in October 2026, with the real size of each 4-bit download, the context window that still fits beside it, the licence and what each model is good at. Every fit is worked out with the same arithmetic as our LLM VRAM calculator.

Open-weight vision language models you can run on your own GPUs, checked against their model cards in October 2026: sizes, licences including Llama 4's EU clause, memory for weights and image tokens, serving support in vLLM, llama.cpp and Ollama, and which model suits photos, charts, screens or video.

Consumer, workstation, unified-memory and data-centre GPUs for running open-weight models on your own hardware, compared from NVIDIA, AMD, Intel and Apple spec sheets on VRAM, memory bandwidth, FP8 and FP4 support, power and list price, with picks by budget and by model size.

Google's Gemma 4 family runs from phone-sized E2B to a 31B dense model, all under Apache 2.0. This guide lists every size and variant, the memory each needs for weights and KV cache, Google's official quantization-aware builds, and the install steps for vLLM, llama.cpp, Ollama, LM Studio and MLX.