Self hosted ai

Which open-weight LLMs actually fit in 8GB of VRAM or RAM in 2026, with measured file sizes, KV cache math from published configs, and Ollama commands for Qwen3.5, Gemma 4, Ministral 3, Granite 4.1, Nemotron 3 Nano, and Phi-4-mini.

A buying guide for running coding agents on your own hardware. Why prompt caching makes generation speed the number that matters, where a cold cache costs you minutes instead, a comparison table of Mac Studio M5 Ultra, RTX 5090, Radeon AI PRO R9700, Strix Halo and DGX Spark with September 2026 prices, and the memory every open-weight coding model actually needs.

How to install Nous Research's open-source Hermes Agent on your own machine and run it at zero token cost on OpenRouter free models - the tool-calling and 64K context requirements, the request limits that actually bite, a free-model fallback chain, and remote access over a tunnel.

Apple's M5 Ultra Mac Studio tops out at 512GB of unified memory at 1.2TB/s. Here is the memory math, which large open-weight LLMs actually fit, measured tokens per second, and three ways to run a model bigger than 4-bit allows: Unsloth GGUF quants, streaming weights off SSD, or a Thunderbolt 5 cluster.

oMLX turns a Mac into an OpenAI-compatible inference server with continuous batching and SSD-backed KV caching. Here's how it works, its new multi-Mac clustering, and how to reach it from your phone or a coding agent with Pinggy.

Fine-tune LLMs locally with Unsloth: QLoRA on a single GPU, Unsloth Studio, Dynamic GGUF quants, and sharing your model over a Pinggy tunnel.

OmniRoute is a free MIT-licensed AI gateway you run yourself: one OpenAI-compatible endpoint in front of 290+ providers and 500+ models. We ran v3.8.48 in Docker, got 99 models resolving with zero configuration, tested combos, compression, MCP, and the CLI, then shared the whole thing over a public HTTPS URL with Pinggy.

Compare the best OpenClaw alternatives in 2026 - Nanobot, NanoClaw, ZeroClaw, Hermes Agent, and TrustClaw. See how each stacks up on performance, sandboxing, security, features, licensing, and self-hosting.

A guide to the best open-weight text-to-speech models you can self-host in 2026, ranked by Artificial Analysis Speech Arena Elo. Compare Breeze TTS 2, Fish Audio S2 Pro, Step Audio EditX, Voxtral TTS, Qwen3-TTS, VoxCPM2, GLM-TTS, Higgs Audio V3, Kokoro 82M and Chatterbox on quality, licensing, hardware and deployment.

A single-file C engine called Colibrì streams GLM-5.2's 744B mixture-of-experts weights off an NVMe drive to run the full model on 25 GB of RAM at 0.05-2 tokens/second. Here's how it works, what Hacker News made of it, and how to check on a queued run from your phone with Pinggy.