SwarmLLM

by enapt · Apache-2.0 · 82 GitHub stars · Experimental · v0.3.276-alpha, Nov 2026

A new desktop app whose users form an internet swarm, or a private group, each holding some layers, with speculative decoding to hide long links.

Mac: Yes
NVIDIA PC: Yes
AMD PC: No
Linux, processor only: Yes
Raspberry Pi and boards: No
Windows: Yes
Android: No
iPhone, iPad: No

Go to SwarmLLM Source code

At a glance

Best for: Sharing GPUs with friends or volunteers over the internet.
Runs on: Windows, Linux (CUDA for RTX 30 or newer, or CPU) and Apple silicon Macs.
How it splits a model: Pipeline, plus speculative decoding.
Connections: The internet, behind home routers.
How devices join: Automatic.
Models: Llama 3.x, Qwen 2.5 and Qwen3 family models.
API: OpenAI and Anthropic compatible.
Rules: Machines more than two releases behind are dropped until they update.

Get started

Linux

Download from the GitHub releases and run ./swarmllm run

Limits to know

Alpha software with forced updates.
Far-apart machines are slow for big models (its own Llama 3.3 70B runs: about 2 tok/s).

We don’t estimate speeds for SwarmLLM.

Sources:

enapt/SwarmLLM on GitHub

All clustering tools

Estimates come from published hardware specs and our open formulas, not from running the model on your machine. Real speed varies with runtime, drivers and settings. Numbers marked “Measured” come from a linked benchmark.