import WorkGraph from './WorkGraph.jsx' const ORG = 'https://huggingface.co/Lets-Fafo' const MODEL = `${ORG}/Qwen3.8-27B-FAFO-W4A16-G32-B70` const questions = [ ['Who', 'Who did the work?'], ['Why', 'Why was it allowed?'], ['Grounding', 'What grounding was used?'], ['Change', 'What changed?'], ['Evidence', 'What evidence exists?'], ['Cost', 'What did it cost?'], ['Trust', 'Can you trust it?'], ] const products = [ { icon: 'agentos', verb: 'Governs the work', name: 'FAFO™ AgentOS', text: 'Authority, evidence, recovery and economics for each unit of work. Self-hosted, on Linux.', href: 'https://letsfafo.com', }, { icon: 'memory', verb: 'Grounds the work', name: 'FAFO™ Memory', text: 'Code, decisions and references, retrieved by meaning and handed to each worker before it acts.', href: 'https://letsfafo.com/fafo-memory.html', }, { icon: 'swarm', verb: 'Runs the work', name: 'FAFO™ Agent Swarm', text: 'Specialized AI agents, governed end to end and replaceable when they fail.', href: 'https://letsfafo.com/agent-swarm.html', }, { icon: 'fabric', verb: 'Powers the work', name: 'FAFO™ Inference Fabric', text: 'A GPU saturation layer for NVIDIA inference.', href: 'https://letsfafo.com/inference-fabric.html', }, ] const Arrow = () => export default function App() { return ( <> Skip to content
LetsFAFO

Lets-Fafo · Hugging Face

You can FAFO.
Your AI shouldn't.

FAFO™ is Fully Autonomous Foundational Orchestration: a self-hosted platform where AI agents complete assigned work inside explicit scope, budgets, independent review and an audit trail. The operator defines intent. AgentOS derives execution.

See what we publish Meet the company behind FAFO
Patent pending · five patents filed NVIDIA Inception Program member Featured on Product Hunt

Why governed work

AI can answer questions. Organizations need work completed.

Ungoverned AI keeps organizations in pilot mode. Autonomous work needs the accountability human teams already have: assigned work, budgets, independent review and an audit trail. A leader should be able to answer these seven questions in seconds.

    {questions.map(([tag, q]) => (
  • {tag}

    {q}

  • ))}

Products

Four parts. One governed loop.

The platform, code, weights and local models run inside your perimeter. Frontier model use is optional and governed.

    {products.map((p) => (
  • {p.verb}

    {p.name}

    {p.text}

    Learn more
  • ))}

What we publish here

Models we adapted and measured on our own hardware.

Each model card says exactly what changed, which hardware and software it was tested on, and what we measured. Nothing beyond that.

Quantized model · Apache-2.0

Qwen3.8-27B-FAFO-W4A16-G32-B70

1,798 tokens per second total throughput from one Intel Arc Pro B70 (32 GB). Qwen3.8-27B reduced from 55.6 GB of BF16 weights to a 20.6 GB W4A16 checkpoint, validated only on the B70 with Intel's XPU build of vLLM.

Open the model card
Total throughput
1,798 tok/s

in + out, 32 clients, prefix caching on

Weights on disk
20.6 GB

from 55.6 GB BF16

Aggregate output
~296 tok/s

generated tokens, sustained, 64 clients

Longest prompt tested
110,000 tokens

single card, no failures

MeetFAFO · public experiment

AI runs the businesses. We publish the results.

AI teams build products, run marketing and serve customers. Humans approve the direction. We publish the revenue, the costs and the lessons.

FAFO's work orchestration and evidence systems, applied to creating and operating real businesses. Status: preparing the first release.

meetfafo.com
) }