The Local Lab.
Free. Private. Yours.
Which mode to run, what your machine can really handle, the models worth the disk space — and the two traps that make people think local is bad.
Choose your mode first.
Local isn't all-or-nothing. Decide how much leaves your machine — the model choice comes next.
Vault
100% local, offline-capable. Client data, contracts, health, unreleased work.
sensitive workConnected
Local handles drafts and heavy lifting; switch to the cloud for the toughest decisions.
★ the everyday strategyCloud-first
Cloud by default; falls back to local only when something can't leave the machine.
low VRAMWhat your can your machine really handle it?
Drag to your machine's memory — on a Mac with Apple Silicon, unified memory counts as VRAM. On a PC, use your GPU's VRAM.
Three jobs. Three champions.
Don't collect models — assign them. Click a command to copy it.
32B-class quality at the speed of a small model. The king of local coding.
ollama pull qwen3-coder:30b-a3b
Frontier chain of thought, distilled. Watch it think before it answers.
ollama pull deepseek-r1:14b
OpenAI's open weights. Chat, tools and agents, in the house style.
ollama pull gpt-oss:20bSix houses. All size.
Every house launches a size ladder — the same brain, scaled up or down. Ollama is the front door to all of them. The options your machine can handle light up (synced to the slider above) — click any dot to copy its pull command.
Qwen3pragmatic · tool-happyDeepSeek-R1thinks out loudGPT-OSSfrontier models · strict
Gemma 3polished · safest
Llamathe default · middle ground
Mistrallean · relaxedAlignment is part of the spec sheet. Gemma and GPT-OSS come locked down from the factory — they refuse anything spicier, which is exactly what you want for family-safe or client-facing use. Llama sits in the middle; Mistral is looser. And there are also fully neutral, which you can steer yourself — that’s Nous Research’s specialty, with open, steerable models built for this kind of control. Choose the right temperament for the right job.
ollama.com
LM Studio (graphical interface)Local is good, for real?
Honest answer: a good local 30B today can do what the frontier could do not so long ago — and for drafts, grunt work, and private documents, that's already most of your tokens.
The pocket-sized brain
Chat, drafts, summaries — quick, light, runs on almost anything.
Yesterday's frontier
Real code, real agents, real context — on your own desk, for free.
Open, but enormous
GLM, DeepSeek, Kimi — open-weight frontier. Too big to house; pennies to rent.
The toughest 5%
Deep reasoning, architecture, good design sense. Send it here — then come back home.
In pure reasoning, the frontier still wins — that's why Connected mode exists. But the advantage of local isn't IQ: it's zero cost per token, zero data outflow, and no rate limits. Route by task, not by loyalty.
Why people think local is bad.
It's usually not the model. It's one of these two settings.
The 4,096-token default
Qwen3-Coder's real context is 256K. Ollama ships it capped at 4K — you're using ~1.6% of the brain, then blaming the model.
OLLAMA_CONTEXT_LENGTH=65536 ollama serveThe KV-cache wall
Context consumes memory: a 70B needs ~42GB just for cache at 128K. Consumer machines top out around 32–64K usable — it's the cache that won't fit, not the model.
q8_0 cuts cache memory almost in half, with minimal quality loss.Three commands to a sealed room.
Ollama exposes an OpenAI-compatible endpoint on localhost:11434 — your agent plugs in like any other brain. Nothing leaves the machine.
# 1 · puxe o coder $ ollama pull qwen3-coder:30b-a3b # 2 · suba o contexto ANTES de julgar $ OLLAMA_CONTEXT_LENGTH=65536 ollama serve # 3 · aponte seu cliente pra localhost $ <seu-cliente-openai-compatible> --model qwen3-coder:30b-a3b --base-url http://localhost:11434/v1
Tonight: drag the slider to your machine, pull the recommended models, and raise the context before judging anything. Your lab is free, private, and always open.