Tools for clustering your devices

Which one to use

• Only Macs: EXO for the easiest setup, or MLX distributed from the command line. Both use Thunderbolt 5 RDMA when the Macs have it.
• A mix of Macs, PCs and boards: llama.cpp RPC, or Mesh LLM if you’d rather join machines with a token than type addresses.
• Several identical Raspberry Pis or small PCs: Distributed Llama, the one tool where more boards really type faster.
• Old laptops and phones on Wi-Fi with too little memory: prima.cpp, or Pooled in a browser, iPhones included.
• A NAS (network attached storage) in the mix: llama.cpp RPC or LocalAI run on it in a container (Docker) and add its memory to the pool. A NAS processor is slow, so it suits holding the model files for the other devices better than doing the work.
• Using Ollama? It can’t cluster; llama.cpp RPC runs the same GGUF files.

15 tools compared

Tool	For	Runs on	Splits the model	Joins by	Licence	Stars	Status
EXOExo Labs	For devices you own	Mac: Yes NVIDIA: Partly Linux CPU: Partly	Pipeline and tensor	Automatic	Apache-2.0	48,222	Active · v1.0.71, Apr 2026
llama.cpp RPCggml-org (llama.cpp)	For devices you own	Mac: Yes NVIDIA: Yes AMD: Yes Linux CPU: Yes Pi: Yes Windows: Yes Android: Partly	Pipeline and tensor	Addresses by hand	MIT	133,581	Active · v0.9.0, Nov 2026
MLX distributedApple (ml-explore)	For devices you own	Mac: Yes NVIDIA: Partly Linux CPU: Partly	Pipeline and tensor	Addresses by hand	MIT	29,079	Active · mlx v0.32.4, mlx-lm v0.32.2, Oct 2026
Mesh LLMMesh-LLM project	For devices you own	Mac: Yes NVIDIA: Yes AMD: Yes Linux CPU: Yes Pi: Partly Windows: Yes	Pipeline	Token or link	Apache-2.0	3,959	Active · v0.78.1, Oct 2026
Distributed LlamaBartłomiej Tadych (b4rtaz)	For devices you own	Mac: Partly NVIDIA: Partly AMD: Partly Linux CPU: Yes Pi: Yes Windows: Yes	Tensor	Addresses by hand	MIT	3,132	Maintained · v0.16.5, Feb 2026
prima.cppZonghang Li et al. (OpenCPIL)	For devices you own	Mac: Yes NVIDIA: Yes AMD: Partly Linux CPU: Yes Pi: Partly Android: Partly	Pipeline	Addresses by hand	MIT	129	Experimental · no releases
PooledNehanth	For devices you own	Mac: Yes NVIDIA: Yes AMD: Partly Windows: Yes Android: Partly iPhone: Yes	Pipeline	Token or link	MIT	628	Experimental · v2.0.0, Oct 2026
ParallaxGradient	For devices you own	Mac: Yes NVIDIA: Yes Windows: Partly	Pipeline	A main machine	Apache-2.0	1,420	Quiet · v0.1.2, Dec 2025
vLLM and SGLang (multi-node)vLLM project, SGLang project	For GPU servers	NVIDIA: Yes AMD: Yes	Pipeline and tensor	Addresses by hand	Apache-2.0	96,479	Active · vLLM v0.33.0, SGLang v0.5.23, Nov 2026
LocalAI (P2P workers)Ettore Di Giacinto and contributors	For devices you own	Mac: Partly NVIDIA: Yes AMD: Yes Linux CPU: Yes Pi: Yes Windows: Partly	Pipeline	Token or link	MIT	50,233	Active · v4.12.1, Oct 2026
GPUStackGPUStack	For GPU servers	NVIDIA: Yes AMD: Yes Windows: Partly	Pipeline and tensor	Token or link	Apache-2.0	5,984	Active · v2.3.0, Oct 2026
PetalsBigScience workshop	Quiet projects	Mac: Partly NVIDIA: Yes AMD: Partly Windows: Partly	Pipeline	Automatic	MIT	10,697	Quiet · v2.2.0, Sep 2023
CakeSimone Margaritelli (evilsocket)	Quiet projects	Mac: Yes NVIDIA: Yes AMD: Partly Linux CPU: Yes Pi: Yes Windows: Partly Android: Partly iPhone: Yes	Pipeline	Automatic	FAIR License (source-available, no business use)	3,133	Quiet · no releases
SwarmLLMenapt	For devices you own	Mac: Yes NVIDIA: Yes Linux CPU: Yes Windows: Yes	Pipeline	Automatic	Apache-2.0	82	Experimental · v0.3.276-alpha, Nov 2026
OllamaOllama	Doesn’t cluster		No	—	MIT	186,147	Active · v0.40.15, Nov 2026

Stars and releases from GitHub on 6 November 2026. “Quiet” means no release or change for months. Also seen but not listed: tools that only send each request to whichever machine has the model, such as LM Studio’s LM Link and NVIDIA’s Personal AI Router, which don’t split a model.