SwarmLLM splits a 27B Qwen model across browser tabs on WebGPU and WebRTC, 408 stars in 18 days

Nehanth Narendrula has released SwarmLLM, a peer-to-peer runtime that splits a 27B language model across browser tabs on different devices and passes the hidden state between them over WebRTC. The repository, published on 1 September and MIT-licensed, reached 408 stars in 18 days. The model it demonstrates is Qwen 3.8 27B, quantized to GGUF Q4_0, which comes to about 15 GB of weights.

The pitch, in the project's own words: "Your phone can't run Qwen 3.8. Your room can." Each device holds a slice of the model's layers. When it is that device's turn, it runs its layers on the local WebGPU adapter, produces a 5,120-float hidden state for the token being generated, and hands it to the next peer over WebRTC. One peer submits the token at the end.

Decode speeds from one GB10 down to a laptop and phone

Narendrula reports decode speeds for the same 27B model across three settings. On a single NVIDIA GB10, the model runs at 9.0 tokens per second plain and 16.1 tokens per second with speculative decoding, and prefills at 44 tokens per second. A MacBook Pro on its own does 6.7 tokens per second plain and 10.8 with speculation. A MacBook Pro paired with an iPhone on the same Wi-Fi, holding 62 and 2 layers respectively, runs at 7.7 tokens per second. Peers across the public internet run the same model at 3.5 to 6 tokens per second.

A recorded demo dated 7 September shows the MacBook and iPhone pair generating 400 tokens at 10.7 tokens per second.

How it differs from WebLLM and MLC

WebLLM and MLC also run models in the browser through WebGPU, and each fits the whole model in a single tab. SwarmLLM shards the model across devices, so the ceiling on model size is the sum of the peers' memory. The tradeoff is a per-token round trip across the network for the hidden state, which is why the internet-scale figures are slower than the single-device ones.

There is no install step. The demo is a URL: open swarmllm.ai/room in a WebGPU-capable browser, share a room code, pick a model, and start. The engine is written from scratch in WGSL for the GPU side and TypeScript for the runtime.

For a team building software, this makes it possible to prototype a mid-sized open model on hardware already at the desk, with no GPU server and no need to download multi-gigabyte weights to every laptop. The cost is latency. A token that walks through two peers waits on the network twice, which fits offline drafting and internal tools where a first-token target is loose.

Source

• Nehanth Narendrula, SwarmLLM repository, README, September 2026.
• SwarmLLM demo.