Hive

Hive turns the Apple devices you already own into one private AI computer. It pools the memory of a household’s Macs, iPhones and iPads so they can run open-weight AI models too large for any one of them, and prompts and answers never leave the home network.

A model too large for a device’s memory can still run on it by streaming its weights from the SSD, but a dense model then reads its whole file from disk for every token, at seconds per token. Hive keeps the model’s layers in the combined memory of the household’s devices instead, so the same model at (nearly) the same quantization runs far faster than the best single device can stream it. A model that already fits on one device runs there, because splitting it would only add network hops.

How it works

• One Mac is the Core. It runs the coordinator, the scheduler, the API gateway and the model store, downloading each model once, checking every file’s digest and splitting it by layer.
• Every device, the Core included, runs a worker that holds a contiguous range of the model’s layers with MLX Swift. Each token’s hidden state passes from stage to stage, and the scheduler places the layers by each device’s measured memory budget and speed.
• Devices pair through the household’s own certificate authority, with no account or cloud, and every link between them is mutual TLS.
• iPhone and iPad take part from the app while iOS allows it. A device that sleeps or leaves is normal: Hive notices within seconds, replans onto the devices that remain, and never presents a broken answer as complete.
• Coding tools and chat clients such as Claude Code and Xcode connect through OpenAI- and Anthropic-compatible endpoints, with a key per client.

Results so far

• Qwen3 32B (4-bit), pooled across a 16 GB MacBook Air (M4), an iPhone 17 Pro and an iPad Pro (M4), decodes about 120 times as fast as the Mac streaming it from its SSD, and prefills a 2,048-token prompt about 14 times as fast.
• Qwen3.8 27B (8-bit), pooled across two MacBook Airs (M5 32 GB and M4 16 GB), decodes about 35 times as fast as the larger Mac alone.
• Claude Code finished a real tool-using task in 5 of 5 runs against Qwen3.5 9B on one Mac.

The model library

Hive offers a short list of open-weight models, chosen and ranked by Hive’s own tests rather than third-party scores. The list is a signed file the app checks for updates, so new models arrive without an app release; the current one is published on the Hive model library page. Weights download from Hugging Face, from the publisher’s own MLX build or else the mlx-community one, pinned to a revision and verified by SHA-256. Hive hosts no weights, and a library change never deletes a model you have installed.

Status

In development. The feasibility work, the runtime core, distribution and recovery (eleven scenarios passed on real devices with the internet unplugged) and the model platform are complete. Now in progress: complete OpenAI and Anthropic API support, and the consumer apps for iPhone, iPad and Mac on the way to TestFlight. Hive will be open source under Apache-2.0 when it reaches TestFlight.