local AI · open models · your hardware or the cloud

mindpool@local:~$ ./build --target=local-ai

Build local AI on your own hardware

Not a course on calling a frontier API. mpl is a toolchain for building local AI — with the curriculum baked in. Run, ground, fine-tune, and ship open models you actually control. No GPU yet? Start on bare-metal cloud you rent by the second; own the hardware once you're ready. Either way you stop outsourcing the three things that decide everything: what it costs, whether it keeps running, and who sees your data.

› see why › read the thesis

› browse curriculum

✓detected Apple M3 Max · 64GB unified — tier T3

✓installed llama.cpp b9873 · mlx-lm · uv (pinned, sha256-verified)

✓pulled Qwen3-8B · Q4_K_M · 4.7GB

✓fine-tune env ready — unsloth + trl

✓enrollment verified on-chain — 0x9f…2aE

✓coach online — gemma4:12b · local ReAct agent, 0 frontier calls

▸module 09 · quantization — unlocked

$ cat ./what-you-actually-learn

The on-device lifecycle, end to end

Every module maps to a real step you run locally — no hand-waving, no "it just works."

01 → 02

Source & run

Read model cards on HF, run GGUF locally, expose a local API.

03 → 05

Prompt, ground & tools

Chat templates and sampling, RAG retrieval, then tool-use agent loops.

06 → 08

Data, tune & evaluate

Curate datasets, LoRA/QLoRA → GGUF with Unsloth, then score on an eval harness.

09 → 10

Optimize & ship

imatrix Q4_K_M quants, serve for throughput, package and hand off.

$ mpl chat

A coach that runs where your models do

mpl chat opens a local AI coach: a ReAct agent on an open model (Gemma 4, Apache 2.0) that answers across the whole curriculum, runs mpl verbs on your behalf — with confirmation — and never calls a frontier API. The agent teaching you to build a local agent is one. Same open weights, same hardware, same loop you'll write in module 05.

It remembers you across sessions — durable memory you approve save-by-save, kept on your disk — cites the actual lessons it teaches from, and follows you from your own hardware to a rented GPU mid-conversation and back, history intact. Swap the model, swap the machine: same coach, same thread.

$ mpl chatone conversation · two machines

you › keep the number 47 in mind for me
coach › Noted.
you › /cloud t4-box
(now: gemma4:12b · http://t4-box:7007 · cloud · serving gemma4:12b)
you › what number did I ask you to keep in mind?
coach › 47.
  coach wants to run save_memory {"text":"Favorite prime: 47."} — allow? [y/N] y
Saved: Favorite prime: 47.  # on your disk — ~/.mpl/coach, never the node's
you › /local
(now: gemma4:12b-mlx · http://127.0.0.1:11434 · local)

› $ mpl chat

$ map ./the-path

Use it. Automate it. Build it.

One path from using AI to building it. Each stage layers real lessons on a real toolchain — start where you are, and go as deep as you want.

useautomatebuild

level 1

Use

You want to get more done with AI today.

curriculum

Run open models on your own machine.

mpl

Install the toolchain and run your first model.

level 2

Automate

where most people are headed

You want to build agents that automate real work.

curriculum

Prompt it, ground it in your data, give it tools. Agent Harness Engineering.

mpl

mpl agent runmpl lab D1…D8mpl chat

level 3

Build

the production track — optional, not required

You want to build the models themselves.

curriculum

Train, fine-tune, quantize, and ship. Inference Engineering · Build a Model from Scratch. Production Local AI planned.

mpl

mpl benchmpl lab A1… / B1…mpl doctormpl burst

› follow the path › see the full curriculum

$ df -h ./hardware

Pick the rig that runs the work

Bandwidth, not capacity, sets your tokens/sec. Cloud-spot is the relief valve for the rare 70B full tune.

› the buyer's guide — 4 tiers