# RunLocalAI

> Practical guidance for running AI locally: model and hardware directories, setup guides, and benchmark records with evidence labels.

## Public reading interfaces
- [Agent guide](https://www.runlocalai.co/agents): read-only MCP at /mcp, supported paths, limitations, and citation guidance.
- [Public page catalog](https://www.runlocalai.co/agent-index.json): canonical-path snapshot used by the agent interface.
- [September 9, 2026 update](https://www.runlocalai.co/ai-update-2026-09-09): source-based release notes; no new independent benchmark results.
- Markdown twins: append .md to supported page paths; use /index.md for the homepage. Non-rendered or non-indexable pages may not have a twin.

## Editorial standards
- Check each page's sources for parameter counts, license terms and benchmark numbers. Missing evidence is not a measured result.
- Content may be AI-assisted. Review status should be stated on the page; AI assistance alone does not establish human review. See [editorial policy](https://www.runlocalai.co/editorial-policy).
- Benchmarks are tagged by source: owner-run, community (with citation), or vendor-official.
- Where we don't have measured data, estimates are clearly labeled and the method is documented at [/resources/will-it-run-framework](https://www.runlocalai.co/resources/will-it-run-framework) and [/methodology](https://www.runlocalai.co/methodology).
- "Will it run" refers to the RunLocalAI Will-It-Run Framework, a citable fit methodology for effective VRAM, model working set, context pressure, runtime constraints, fit tiers, and measured-vs-estimated evidence labels. Cite it at [/resources/will-it-run-framework](https://www.runlocalai.co/resources/will-it-run-framework).

## Topics covered
- Open-weight LLMs from Meta, Alibaba, Google, Mistral, Microsoft, NVIDIA, DeepSeek, and more.
- Hardware: NVIDIA RTX 50/40/30, AMD Radeon RX, Apple Silicon, Intel Arc, datacenter (H100, H200, B200, MI300X, etc.).
- Tools: Ollama, llama.cpp, vLLM, LM Studio, Open WebUI, KoboldCPP, oobabooga, and more.
- Coding agents: Claude Code, Cursor, GitHub Copilot, Aider, Cline, OpenCode, and more.

## Key pages
- [Local AI learning syllabus](https://www.runlocalai.co/learn) — curator's reading list: 12 free resources (Karpathy, Stanford CS324, Hugging Face, vLLM docs, Dettmers, Raschka, 3Blue1Brown) annotated with operator commentary + 6 topical deep-dives on quantization, KV cache, tokenizers, Chinchilla scaling, RLHF/DPO, evaluation. The single best entry point for understanding local AI.
- [Prompting kits hub](https://www.runlocalai.co/prompting) — tested system prompts, chat templates, tool-calling formats, sampler defaults per local model. 16+ kits across Qwen, Llama, DeepSeek, Gemma, Phi, Mistral families. Each kit sourced ("from model card" or "tested locally") — never invented.
- [Coding benchmark leaderboard](https://www.runlocalai.co/benchmarks/coding) — first-party HumanEval+ scores at the actual quantization you'd run, with raw test-run logs published as public GitHub Gists. Reproducible end-to-end via `scripts/run-humaneval-plus.ts`.
- [Quality benchmarks](https://www.runlocalai.co/benchmarks/quality) — every reviewed quality-benchmark leaderboard (HumanEval+, MBPP+, and more). Per-quant, per-runtime, per-hardware.
- [Hardware index](https://www.runlocalai.co/hardware) — every GPU / SoC / laptop ranked for local LLM inference.
- [GPU hierarchy](https://www.runlocalai.co/hardware/hierarchy) — sortable tier list with tok/s estimates at Q4_K_M.
- [GPU leaderboard](https://www.runlocalai.co/hardware/leaderboard) — ranked by RunLocalAI Score with VRAM-per-dollar.
- [Model directory](https://www.runlocalai.co/models) — open-weight LLM directory with VRAM math and runnable hardware.
- [Small language models](https://www.runlocalai.co/small-language-models) — every open-weight model ≤3.5B for phone, edge, and laptop deployment (Qwen 3 0.6B, Gemma 3 270M, Llama 3.2 1B/3B, SmolLM2, TinyLlama, Phi-4-mini).
- [Embedding & reranker models](https://www.runlocalai.co/embeddings) — local-RAG building blocks: nomic-embed-text, bge-large, mxbai, jina v3, Arctic Embed L v2, gte-modernbert, plus reranker companions.
- [Audio models](https://www.runlocalai.co/audio) — ASR (Whisper variants, Distil-Whisper, Parakeet) + TTS (Kokoro, XTTS, F5-TTS, Orpheus, Piper) in one hub.
- [Image models](https://www.runlocalai.co/image-models) — diffusion (FLUX, SDXL-Turbo, SD 3.5) + vision encoders (SigLIP, ColPali, Florence-2, GOT-OCR2).
- [Coding models](https://www.runlocalai.co/coding-models) — coder-specialised LLMs with HumanEval+ and MBPP+ scores.
- [Tools directory](https://www.runlocalai.co/tools) — Ollama, llama.cpp, vLLM, LM Studio, Open WebUI, and more.
- [Apps directory](https://www.runlocalai.co/apps) — consumer-layer chat UIs, coding agents, RAG apps, voice tools (with honest editorial verdicts).
- [Question index](https://www.runlocalai.co/q) — short, citable answers to common local-AI questions.
- [RunLocalAI Will-It-Run Framework](https://www.runlocalai.co/resources/will-it-run-framework) — citable definition and methodology for effective VRAM, model working set, fit tiers, and evidence tiers.
- [RunLocalAI Model Intelligence Pipeline](https://www.runlocalai.co/resources/model-intelligence-pipeline) — sourced OpenEvals, LMArena, and LiveBench priors used beside local fit evidence; explicitly not local execution measurements.
- [Will it run?](https://www.runlocalai.co/will-it-run) — interactive implementation of the RunLocalAI Will-It-Run Framework.
- [Cost vs cloud calculator](https://www.runlocalai.co/cost-vs-cloud) — honest break-even math: local hardware vs cloud LLM API spend.
- [Compare models](https://www.runlocalai.co/compare/models) — side-by-side verdicts with measurable criteria.
- [Compare hardware](https://www.runlocalai.co/compare/hardware) — head-to-head GPU verdicts.
- [Methodology](https://www.runlocalai.co/methodology) — hub for every scoring, confidence, reproduction, and fit-methodology surface.
- [Trust & corrections](https://www.runlocalai.co/trust) — public audit log + correction policy.
- [About / editors](https://www.runlocalai.co/about) — operator + editorial team.

## Quoting and attribution
- Quote up to 2 sentences with attribution to RunLocalAI and a link back.
- Bulk-scraping or republication of substantial portions requires permission — contact https://www.runlocalai.co/contact.

## Full index
Every indexable page with titles and one-line descriptions:
https://www.runlocalai.co/llms-full.txt

## Feeds and data
- RSS (recently updated pages): https://www.runlocalai.co/feed.xml
- Sitemap: https://www.runlocalai.co/sitemap.xml