🐝 SMOL SWARM

A colony of tiny models that outperform giants.

🚨 The Problem

Building local AI agents today is broken:

Running four 70B models for a multi-agent pipeline requires $40k in GPUs.
Most small models (<2B) fail at structured JSON, tool-calling, and coordination.
There is no shared protocol between micro-models, so every developer rebuilds glue code from scratch.
The open-source ecosystem has thousands of chat models but zero coordinated swarms.

✅ The Solution

SmolSwarm is a complete ecosystem of sub-2B micro-models designed to work together as specialized workers in local multi-agent systems.

Benefit	Detail
💾 Tiny Footprint	Full 9-model swarm runs in ~3.5 GB RAM (Q4_K_M)
⚡ Blazing Fast	100–500+ tok/s per agent on a laptop CPU
🔗 Plug & Play	Unified JSON envelope protocol — chain any agent to any agent
🏠 100% Local	No API keys. No cloud. No data leaving your machine
🆓 Open Source	Apache 2.0 — use commercially, modify freely

📡 The Unified Swarm Protocol

Every model in the swarm speaks the same JSON envelope. This is what makes SmolSwarm a system, not just a collection of models.

Request Format

{
  "task_id": "uuid-1234",
  "task": "extract",
  "input": "Hi, I'm Alex, 29, living in Berlin.",
  "schema": {"name": "string", "age": "number", "city": "string"},
  "metadata": {"source_agent": "router", "priority": "high"}
}