Comparison

How Pooled differs from other projects, and when to pick one of them instead. Other projects split models across machines, or run models in a browser tab. Pooled does both: it splits one model across devices, and every device runs its share in a browser tab.

Not a ranking

Each published speed below is for a different model on different hardware, taken from each project’s own sources. Read it as what each project can do, not as a head-to-head result.

At a glance

Project	Runs in	Pools several devices	Install per device	Coding agent built in
Pooled	Browser tab (WebGPU)	Yes: layers split across laptops, desktops and phones over WebRTC	None: open a URL	Yes: Code mode runs in the tab
exo	Native (MLX); macOS app, Linux on CPU for now	Yes, with RDMA over Thunderbolt 5 on Macs	macOS app or build from source	No
llama.cpp RPC	Native	Yes: remote ggml devices	Build with -DGGML_RPC=ON on every host	No
distributed-llama	Native (CPU, experimental Vulkan)	Yes: tensor parallel, power-of-2 node counts	C++ build plus a Python launcher	No
Petals	Native Python, public or private swarm	Yes: layers served by volunteers	pip package, PyTorch	No
Mesh LLM	Native (Rust, llama.cpp) plus a web console	Yes: layer splits for dense models, expert sharding for MoE, over QUIC	Install script, Homebrew or distro package	No own agent; launches outside agents against the mesh
WebLLM	Browser tab (WebGPU)	No: the whole model in one tab	None for users; npm or CDN for developers	No
Transformers.js	Browser or Node (ONNX Runtime, WASM or WebGPU)	No	None for users; npm or CDN for developers	No

Published speeds

Project	Published speed	Hardware	Source
Pooled	35B MoE: 40–46 tok/s plain, 65–84 speculative on one device; 12.9 / 20.6 on two devices at 20 ms emulated latency	One GB10 (DGX Spark)	Benchmarks
exo	31.9 tok/s Qwen3 235B; 32.5 tok/s DeepSeek V3.1 671B	4× M3 Ultra Mac Studio	Jeff Geerling
llama.cpp RPC	Qwen3 235B: 20.4 tok/s on 1 node, 15.2 tok/s on 4 nodes	4× M3 Ultra	Jeff Geerling
distributed-llama	Llama 3.2 3B Q40: 5.95 tok/s on 1 Pi 5, 13.68 on 4; Qwen3 30B A3B: 13.04 tok/s on 4	Raspberry Pi 5	#165, #255
Petals	Up to 6 tok/s Llama 2 70B; up to 4 tok/s Falcon 180B	Public swarm	README (last commit 2024-08-25)
Mesh LLM	None published in the README
WebLLM	41.1 tok/s Llama 3.1 8B, against 57.7 native MLC-LLM	M3 Max	paper
Transformers.js	WebGPU up to 100× faster than WASM		v3 post

What is different about Pooled

• Nothing to install on any device. Every device opens a URL. Of the projects that pool devices, the others all need a native install or a build on each machine.
• Phones can join. A phone can ask questions, and hold layers when the computers can’t hold the model.
• Direct links between browsers. Model traffic goes device to device over WebRTC. No server does inference, and there is no account.
• Our own engine. Pooled’s kernels are hand-written WGSL, not WebLLM, MLC or llama.cpp. That is what lets it split a model mid-way and run hybrid DeltaNet models. See Engine and kernels.
• A coding agent in the tab. Code mode writes, runs and fixes a small web app on the room’s model. The room can also serve its model to outside tools through the Serve API.

When to pick something else

• You need the fastest single machine. Native llama.cpp is still about 2× faster on plain decode and much faster on prefill. See Benchmarks.
• You have several Macs with Thunderbolt 5. exo’s RDMA links are far faster than Wi-Fi, and it runs much bigger models.
• You want one model in one tab, embedded in your own app. WebLLM and Transformers.js are libraries built for that.
• You want to share a model with strangers. Pooled rooms are for people you trust. See Security model.

Related projects

Besides the projects in the table, these coding agents work well with local models: pi, little-coder (built on pi) and aider.

Model weights and the tokenizer come from Qwen, hosting from Hugging Face, and signaling from PeerJS. llama.cpp’s speculative-decoding graph for Qwen 3.5 was the reference for Pooled’s.