ABONVERO LEARN · LOCAL AI AND AGENTS
Run AI on your own machine,
and know what to expect.
A guided course on local language models and the agents built on them: what fits on your graphics card, why context length costs speed, why agents need tools, and how to run them without surprises. Every number comes from our own three-machine test fleet. The first 1 lesson is free.
- Ollama
- LM Studio
- llama.cpp
- Agent frameworks
LESSON 1 · PRIVACY, COST AND LIMITS
Why run AI on your own machine
A local model runs entirely on your computer through a runner such as Ollama, LM Studio or llama.cpp. Nothing you type leaves the machine and there is no per-message bill. The trade-off: models that fit on a desktop graphics card are smaller and weaker than the big hosted ones, and you supply the hardware, electricity and maintenance.
1/2 Which job is the best fit for a small local model?
Why this course is different
Local AI can’t run inside a web page, so this is a guided course rather than a simulator: a short explanation, then questions that check you understood it, each with the reasoning behind every answer. What makes it worth reading is where the numbers come from. In September 2026 we ran three machines continuously for two days: an RTX 3080, a GTX 1070 and an Apple M1. We measured every model on every card and kept a record of what the agents actually delivered. The course teaches from those results, including the failures.
What we measured
Tokens per second is roughly how fast text appears. “On GPU” is how much of the model fitted in graphics memory; below 100%, the rest spills into slower system RAM.
| Hardware | Model and context | Tokens/s | On GPU |
|---|---|---|---|
| RTX 3080 · 10 GB | 12B model, 4k context | 65.1 | 100% |
| RTX 3080 · 10 GB | 12B model, 16k context | 26.4 | 88.5% |
| GTX 1070 · 8 GB | 9B model, 80k context | 26.8 | 100% |
| GTX 1070 · 8 GB | 27B model at 2-bit, 16k | 2.3 | 69.3% |
| Apple M1 · 16 GB | 12B model (MLX), 4k | 7.4 | 100% |
Two lessons stand out. First, bigger wasn’t better on fixed hardware: the 27B model was the slowest on every card that could hold it. Second, context isn’t free: the same 12B model ran 2.5 times slower at 16k context than at 4k.
Agents: tools beat size
We ran agents through an agent gateway (OpenClaw) in two ways. Workers forbidden from using tools produced proposals and very little that could ship. One small 9B model allowed to run the tests itself and read the real errors went from 4 of 11 to 8 of 11 passing, the only setup that kept improving. If you build local agents, give them a way to check their own work, then supervise what they deliver rather than how busy they look.
What you’ll learn
- Why run AI on your own machinePrivacy, cost and limitsFree
- Will the model fit? Parameters, memory and quantisationChoosing a model for your hardwarePro
- Context length isn’t freeSpeed versus how much it can readPro
- Local agents need tools, not just a bigger modelAgent design that worksPro
- Running agents safely on your own machinesSupervision, permissions and shutdownPro
Measurements: Abonvero test fleet, September 2026, Ollama runner; your results will vary with drivers, model builds and settings. Tool names are examples, not endorsements. Operated by Andronykus Technologies LLC · contact@abonvero.com