Mac · Apple Silicon · Windows · x64
0.6GBProvider guide
Serve a model on Braise.
This guide is for people who want to share a machine. To use models, you only need the Braise app: no GPU, runtime or local model.
Which runtime do you use?
Pick the one that already serves your model.
llama.cpp
Runs GGUF models, with a built-in local HTTP server.
InstallOpens in a new tabSupported models
Supported modelsPlatformSizeActions
Mac · Apple Silicon
5.7GBMac · Apple Silicon · Windows · x64
19GBOllama
Runs models locally, with a command-line tool and an HTTP API.
InstallOpens in a new tabSupported models
Supported modelsPlatformSizeActions
Mac · Apple Silicon · Windows · x64
—ollama pull qwen3.5:0.8bMac · Apple Silicon
6.6GBollama pull qwen3.5:9bMac · Apple Silicon · Windows · x64
17.7GBollama pull qwen3.8:27boMLX
Runs MLX models on Apple Silicon Macs.
InstallOpens in a new tabSupported models
Supported modelsPlatformSizeActions
Mac · Apple Silicon
0.7GBMac · Apple Silicon
6.3GBMac · Apple Silicon
17.9GBKnown beta limitations
Two runtime issues can affect a shared machine during the public beta. Here’s what happens and what to do.
Ollama
Sharing stops after two quiet hours
Braise asks Ollama to keep the shared model in memory for two hours after its last request. If nobody uses your machine’s model for longer than that, Ollama unloads it. During the beta, Braise may not load it again by itself, and your machine then quietly stops sharing.
To share again, restart Braise, or switch sharing off and back on. Setting OLLAMA_KEEP_ALIVE doesn’t help: Braise’s two-hour setting takes priority over it.
We’ll fix this in Braise after the beta launch.
oMLX
Unfinished thinking can show up as the answer
When a request with reasoning (“thinking”) turned on reaches its maximum output length before the model finishes thinking, oMLX 0.6.x returns the unfinished thinking as the visible answer instead of as reasoning.
This comes from oMLX and is already fixed upstream in oMLX 0.7.0rc1 (pull request jundot/omlx#3809). Until Braise supports a stable oMLX release with the fix, give requests a larger maximum output length.
Want to share a machine?
Put a machine you already own to work.
Braise isn’t available to download yet. Leave your email to hear when the free public beta opens. Providers will earn temporary Compute Credits, not money.
Tell us about your hardware, your runtime and what you’d use credits for. Beta credits can’t be transferred, cashed out or converted, and balances reset before paid V1. Electricity and hardware costs are yours, and there’s no guaranteed volume of jobs.
Email me when the beta opensBeta opening
Get notified when the beta opens.
Tell us if you want to use models, share a machine, or both. We’ll email you when the free public beta opens. This doesn’t create an account or reserve a place.
The beta uses promotional Compute Credits, which are not money. They can’t be bought, transferred, cashed out or converted, and all beta balances reset before paid V1.