Are You Loco Enough?

LocoLab is not for everyone. It is for people who want to understand what local AI can actually do — and are willing to get their hands dirty finding out.

If any of these sound like you, welcome to the lab.

💰 The Budget Rebel

You refuse to pay per-token for something your own hardware can do. You’ve done the math on API costs and it offends you. You’d rather spend a weekend setting up a local stack than sign up for another subscription you’ll resent.

Your entry points:

LocoPuente — a full local AI service stack on hardware you own and control
LocoLLM — a routed specialist model that runs free on consumer GPUs
AI Landscape — honest comparison of local vs cloud options, including cheap API paths

And the bridge is built for you specifically — frontier-equivalent tools without frontier costs is the bridge’s whole job. See The Local AI Opportunity for the longer argument: why the meter is unsustainable, and what that opens up.

🤖 The Tinkerer

You want to understand how LLMs actually work by cracking them open and rewiring the internals. Reading about fine-tuning is not enough. You want to train a real adapter, measure whether it helped, and understand why.

Your entry points:

LocoLLM — adapter training, evaluation harnesses, and a router you can improve
LocoBench — systematic benchmarking infrastructure to measure what you built
Getting Started — technical foundations: inference, VRAM, quantisation, the full stack

The harness is the part you’ll most enjoy taking apart — see the thesis for why the engineering around the model matters more than the model.

🔬 The Researcher

You need reproducible local inference for experiments. You want to test whether specialist routing actually beats a generalist on scoped tasks — and publish honest results either way. You are not interested in vibes.

Your entry points:

LocoBench — VRAM-tier benchmarking with real hardware, real cards, honest baselines
LocoConvoy — multi-GPU parallelism experiments on consumer PCIe hardware
LocoAgente — agentic scaffolding research: can small models think in loops?
Research — active and planned studies across the lab

Our methodology is laid out in the thesis: honest baselines, surfaced uncertainty, status markers on every claim. And findings shows you what’s measured, what’s claimed, and what would invalidate each.

🏫 The Educator

You teach AI, computing, or a professional discipline and want a real project your students can contribute to. Not a toy demo. Real infrastructure that grows with every cohort. You also want rehearsal environments where students practise professional skills before they face the real thing.

Your entry points:

LocoEnsayo — AI-populated rehearsal environments: security audits, requirements gathering, difficult conversations
LocoLLM — a teaching and research framework students build by contributing adapters, benchmarks, and routing improvements
Why Local AI — the case for local AI in education and institutional contexts

The “conversation, not delegation” principle is the pedagogical heart of the lab — it’s what the Cognitive Strategy Transfer and Keep Asking research threads are about, and why the rehearsal environments are designed the way they are. For the future-of-careers angle, The Local AI Opportunity lays out the small-business and consulting openings the next generation of operators is walking into.

🔐 The Vault

Your data does not leave your machine. Period. Medical notes, legal research, personal journals, proprietary code, student assessment work — local inference is not a convenience, it is the only acceptable path. You do not need to be convinced. You need the stack to work.

Your entry points:

LocoPuente — local AI services for institutions and individuals who cannot or will not use cloud inference
AI Landscape — why “private by policy” is not the same as “private by architecture”
Why Local AI — data sovereignty, compliance, and the structural argument for local inference

Local AI is not just “private by policy” but private by architecture — the bridge synthesises capability and privacy in a single stack you own.

⚙ The Scrapper

You know the best gear does not make the best work. A $150 secondhand GPU and sharp training data might just surprise you. You are assembling capability from what is available, and you want to know exactly where the floor is.

Your entry points:

LocoBench — floor-representative benchmarks: worst card per VRAM tier, honest baselines
GPU Inventory — the actual secondhand fleet running these experiments
Economics of Local Training — what local AI actually costs to build and run

The whole “engineer before hardware” principle was built for you. Five llama.cpp flags + system RAM letting an eight-year-old GTX 1060 run a 30-billion-parameter MoE model is the kind of finding the floor produces. And once you know where the floor is, The Local AI Opportunity maps out who needs that capability — small accounting firms, regional legal practices, local consultancies — and what the work of building it for them actually looks like.

🏢 The Operator

You run a small or medium organisation — a clinic, a firm, a consultancy, a school department. You need AI tools that work for your team without sending client data to someone else’s machine, and without an enterprise licence that costs more than your hardware. You are not interested in the research framing; you need pragmatic, reliable, accountable infrastructure the people in your office can actually use.

Your entry points:

LocoPuente — a full local AI service stack deployable on a server in your office and accessed from any device on your LAN
The Local AI Opportunity — the strategic argument: who benefits from local AI, why now, and what the work of building it for them looks like
AI Landscape — honest comparison of local vs cloud, including the cases where cloud is the right call

The “bridge experience, locally” principle is built specifically for the situation you are in: capability and governance in one stack you own. See Why Local AI for the structural argument you can hand to a partner, an auditor, or a board.

🛠️ The Builder

You ship products. Local AI is a backend in your own software — an app, a service, a tool you are putting into people’s hands. You do not need to train models from scratch; you need a stable inference layer, adapters that fit your domain, and a harness that keeps working when you turn around. Research-grade is overkill, toy demos are insulting, and what you want is something between the two that does not break in production.

Your entry points:

LocoLLM — the routed specialist framework treated as a backend SDK: adapters as your domain layer, the router as your dispatch
LocoAgente — the multi-turn harness as a building block: orchestration patterns you can compose into your own workflows
LocoPuente — deployment patterns worth borrowing from, even if you ship something different

The “specialize and harness” principle is your north star — the engineering around the model is your product surface, and that is the part frontier providers will not give you. The LocoLLM-as-LocoAgente-inference integration is the canonical pattern: compose the lab’s projects into one production stack rather than wiring everything yourself.

📋 The Skeptic

You have not decided local AI is ready. You have heard the hype, you have heard the dismissals, and you want neither. You want stress-tested claims, honest negative results, and an evaluator-friendly path through what is measured versus what is still hoped for. You expect to find the floor lower than the marketing suggests — and you would rather find that out from a lab that publishes negative results than from your own production incident.

Your entry points:

Findings — what is measured, what is claimed, what is unmeasured, and what would invalidate each claim
LocoBench — honest baselines on real hardware, including the worst card per VRAM tier
AI Landscape — the comparison page that names the cases where cloud beats local

The “map the floor honestly” methodology is the answer to your scepticism, not a deflection from it: status markers on every claim, invalidation conditions stated up front, and a willingness to publish negative results. If you find the lab overclaiming, that is the page to push back from. We would rather hear it.

Not sure where to start?

If you are new to local AI entirely, Getting Started covers the technical foundations without assuming prior knowledge. If you want to understand the broader landscape before committing to anything, AI Landscape gives an honest comparison of every option — including the ones that are better than LocoLab for your use case.

Loco by name. Serious by intent.