field notes
An LLM gateway is a single OpenAI-compatible endpoint that routes requests across model backends — your own GPUs and the cloud. Here's what a gateway (a.k.a. LLM router) actually does, and why unifying local + cloud behind one URL is the setup serious local-LLM users land on.
read the guide →
An honest, ranked guide to the best open-weight LLMs you can run on your own GPU in 2026 — Llama, Qwen 3, DeepSeek, Mistral, Gemma, Phi — each with size, VRAM, and what it's actually good at. Plus the one tool that tells you whether your card can run it.
read the guide →
The best open-weight coding models you can run on your own GPU in 2026 — Qwen3 Coder, DeepSeek Coder, Codestral/Devstral, and more — ranked by size and VRAM, plus how an OpenAI-compatible gateway lets Cline, Aider, Continue, and other agents use them.
read the guide →
Can you run Claude Code on a local LLM? Technically yes — but there are two real catches: it speaks the Anthropic API, and its large system prompt eats a small model's context window. Here's the honest path, plus the lighter terminal agent that actually works well on your own GPU.
read the guide →
Ollama is the default on-ramp to local LLMs, but it isn't the only one. An honest rundown of the best alternatives — LM Studio, Jan, llama.cpp, LocalAI, vLLM — and where a gateway like Wide Area Intelligence fits when localhost stops being enough.
read the guide →
LM Studio has the best desktop GUI for local LLMs, but it's closed source and single-machine. The best alternatives in 2026 — Jan, Ollama, llama.cpp, GPT4All, vLLM — and where a gateway like Wide Area Intelligence makes any of them reachable from anywhere.
read the guide →
OpenRouter unifies cloud LLM providers behind one API — but every token still bills to someone else's GPU. The best alternatives in 2026, including the option OpenRouter doesn't offer: routing to GPUs you own first, with cloud as failover.
read the guide →
A practical 2026 guide to running Meta's Llama 3 family on your own GPU — picking the right size and quant, the fastest way to get it chatting, and how to make it reachable from your other machines.
read the guide →
Alibaba's Qwen3 family — including the standout Qwen3-Coder models — is some of the best open-weights you can run at home. A 2026 guide to picking a size, running it on your GPU, and making it reachable from your editor anywhere.
read the guide →
DeepSeek's R1 reasoning models are open-weight — but the full model is enormous. The 2026 guide to running DeepSeek locally: which distill actually fits your GPU, how to run it, and how to reach for the full model only when you need it.
read the guide →
Your code calls one OpenAI-compatible endpoint, but each request quietly walks three stages — edge cache, your own GPU, then capability-aware cloud failover. Here's the whole decision path, and the settings that shape it.
read the guide →
Swapping models on the fly — for cost or for resilience — quietly breaks any request that needs vision or tools, if the substitute can't do them. Here's how the gateway routes by what a request actually needs, and how to set a vision model and backups.
read the guide →
Point can1357's oh-my-pi (omp) terminal agent at hardware you own: install it, declare a WideAreaAI provider in models.yml, set it as the default model, and code against your own Gemma node with zero token costs.
read the guide →
Aider is a terminal AI pair programmer that edits your code and commits the diffs. Point it at a coder model running on your own GPU through Wide Area Intelligence and you get git-aware AI pairing with no per-token bill, no rate limits, and full privacy.
read the guide →
Continue is the open-source Copilot alternative — chat, autocomplete, and inline edit in VS Code and JetBrains. Wire it to a Qwen Coder model on your own GPU and pay nothing per request, with no code leaving hardware you own.
read the guide →
Cline is an autonomous VS Code agent built for frontier models — so most local LLM setups disappoint. Here is what actually works: a 32B coder model, a big context window, and your own GPU behind one endpoint.
read the guide →
A no-fluff cost breakdown: real 2026 API prices, real electricity math for an RTX 4090, honest break-even points, and the hidden factors a calculator won't show you. With a verdict for every usage level.
read the guide →
Stop guessing whether a 70B model fits your GPU. The real formula is weights + KV cache + overhead — here are the numbers for every popular model, every context length, and every GPU tier, with no marketing rounding.
read the guide →
Open a GGUF repo on Hugging Face and you get 20 files. This is how the naming works, how much quality each quant really loses, why smaller is faster, and which one to actually download.
read the guide →
Four ways to run the same open-weights model on your own hardware — and why they're different jobs, not competitors. An honest breakdown of where each one wins, with a full feature comparison.
read the guide →
Your model runs great on localhost — until you're at work, on your phone, or shipping an app. A clear-eyed comparison of port forwarding, Tailscale, ngrok, and Cloudflare Tunnel, plus the auth architecture that makes a public endpoint safe.
read the guide →
A coding agent uses your GPU for two hours a day. The other twenty-two, $1,600 of silicon sits idle. Here's how to queue overnight batch jobs — dataset labeling, embeddings, summarization — and let your nodes chew through them while you sleep.
read the guide →
Point Alibaba's open-source coding agent at hardware you own: deploy a Qwen Coder model to a node, wire up three environment variables, and code with zero token costs and full privacy.
read the guide →