Self-learning inference
Using frontier models to teach your local models what to do.
Lakuna answers on your machine first and checks the result against a quality gate. Only what fails escalates to the cloud, and every escalation returns a reusable strategy your local model uses next time.
Fig. 1 · one requestlakuna serve
Local model
Playbook-augmented, tried first
Confidence check
pluggablepass
Returned locally, free.
fail
Escalate, extract playbook entry.
$ npm install -g lakuna && lakuna init
Why Lakuna
The short version.
The problem
A model you pick yourself never improves. Every time it struggles, that experience just makes some cloud company's AI smarter, not yours.
Free by default
Your computer tries to answer first. It only costs anything on the rare times it needs help from the cloud.
Learns from its mistakes
Every time it needs help, it remembers the answer - so next time, it doesn't need to ask again.
Double-checked, not guessed
A simple, honest check decides if an answer is good enough before it's returned to you - never a black box.
Turns real use into a smarter model
Turn weeks of real questions and answers into an upgrade for your own AI, whenever you're ready.
Works with what you already run
Ollama, llama.cpp, vLLM, LM Studio, or Jan.ai - Lakuna works with the tools you've already set up.
No middleman
There's no account key to hand over to us. The only key involved is the one you already have with your own AI provider.
The tradeoff
Local is free and private. Cloud is capable. You shouldn't have to choose.
Local-only inference is cheap and stays on your hardware, but the accuracy gap makes it impractical for real work. Lakuna makes that gap shrink automatically, one escalation at a time.
Local only
Zero cost. Nothing leaves the machine. Fast. But not always good enough, and no way to improve without retraining.
Cloud only
Capable and consistent. But it costs money on every call, adds latency, and every prompt leaves your hardware.
Setup & upgrades
Starts smart, gets smarter.
Most people just pick a local AI model and hope for the best - and it never improves. Lakuna looks at your computer, recommends a model that actually fits it, and keeps a record of everything it had to ask for help with, so that turns into a real upgrade over time.
- One command checks your computer's memory and processing power, then suggests one model that fits - not a confusing list of a hundred options
- Every time it needs help, it keeps a receipt right on your machine - what happened, what it cost, how long it took
- Those receipts turn into short lessons the local model reuses on the next similar question
$ lakuna init
Recommended local model: qwen2.5:14b-instruct-q4_K_M
Wrote configuration to lakuna.yaml
$ lakuna check
config OK: lakuna.yaml - 2 model(s)
$ lakuna serve
Proxy listening. Point your app at it and go.Runs on your machine
Its own dashboard. Nothing leaves your computer.
The router and the checks run on your machine, and so does the dashboard - local vs. cloud volume, what you've spent, and what you've saved, without any of it going through us.
Requests
12,840
Cost spent
$58
Cost saved
$2,310
For developers
One address. Nothing to wire up yourself.
Point your app at Lakuna's address instead of the provider's - nothing else about your code changes. It tries your local model first, checks the answer, and calls your own frontier key only if it needs to.
POST /v1/chat/completions
POST /v1/messages
-> answered locally, or escalated to your
own frontier key if it had to
-> a receipt either way: cost, latency, tierworks with the OpenAI and Anthropic client libraries you already use, unmodified
No account key to hand over to us - the only key in this picture is the one you already have with your own provider.
FAQ
Plain answers before you install anything.
Does the cloud ever see what I type, if my local AI answers fine?
No. It always tries on your computer first. Your question and its answer only leave your machine if that first try isn't good enough - and even then, only once.
What local AI tools does this work with?
Ollama, llama.cpp, vLLM, LM Studio, and Jan.ai today - all running directly on your own machine, not over the internet.
Do I need a fancy graphics card?
No. One command checks what your computer actually has and recommends a model sized to fit it, not a confusing list of a hundred options.
How is this different from just using a cheaper local AI model?
A model you pick yourself never improves - every time it struggles, that experience just makes some cloud company's AI smarter, not yours. Lakuna keeps a record of every time it needed help, so that experience can make your own AI better instead.
Will I need to change how my app talks to AI models?
No - point your app at Lakuna's address instead of the provider's, and nothing else changes. It tries your local model first and only calls out, on your own account, if it needs to.
What happens to the questions it sends to the cloud?
It keeps a receipt of what happened, right on your machine, and remembers a short lesson from it so it can reuse that next time. Nothing is sent to us.
Is this free to start?
Yes - it's open source. Install it and point it at the local AI you already have running. The docs walk through the whole thing.
Using frontier models to teach your local models what to do.
Stop letting that go to waste. Install it, point it at the AI you already run, and let every hard question make it smarter.