Open source · Apache-2.0 · Runs on your computer
Open models, fitted to your rig.
Rigbrew scans your computer, shows which open AI models can actually run on it, and runs a build that was measured for that class of hardware. Chat and pictures share one conversation today; video and voice download now and join it next. No usage limits, and nothing you type goes to a model provider.
macOS beta comingWindows next, Linux later
Download status and requirements01 Will it fit
Pick a machine. Watch what fits.
Rigbrew never hands a model all of the memory. It keeps room for the system and your other programs, then seats each model in what is left, and picks the lightest version that still clears its quality bar.
may hold a model, of 16 GB unified memory, shared by the processor and the GPU.
5.3 GB stays with the system and your other programs.
Class Apple Silicon, up to 16 GB
- Steep 4B Chat Balanced · q54.1 GBFits on the GPU
- Crema 6B Pictures Balanced · int89.7 GBFits on the GPU
- Pour 5B Video smallest version: q411.3 GBToo big 0.6 GB short
Download only for now: running it in Rigbrew comes later.
- Whistle 82M Voice Standard · f320.7 GBFits on the processor
Download only for now: running it in Rigbrew comes later.
Chat and pictures run here today. Video needs 0.6 GB more than this machine can give.
Memory figures are the manifest’s estimates. The rules are core/src/hardware/tier.rs and core/src/select.rs, run here in your browser. One model is in memory at a time, so each is measured against the whole budget.
02 One chat, every model
Ask for a picture in the middle of a conversation.
Each family installs on its own and the chat uses whatever is installed. Ask Steep for a picture and it calls Crema; switch the box to Picture and you go straight there. Video and voice download today and join the chat later.
-
You ask, in plain words.
“Draw a paper boat on a pond, then say one sentence about it.”
-
Steep asks for a picture.
The chat model calls generate_image by itself. No mode to switch, no second app.
-
Crema draws it, here.
1024 × 1024, 8 steps, 50.5 s on the GPU of a MacBook Pro M5 Max.
-
Steep finishes the turn.
13 tokens at 113.3 tokens per second, measured by the app in this chat.
-
One model in memory at a time.
The bay along the bottom shows what is loaded. Rigbrew swaps models as the chat needs them.


03 Measured, not promised
Every build is checked against the original.
Before a build ships it is compared with the full-precision model it came from. The dashed line is the bar; Rigbrew takes the lightest build above it, not the biggest one that fits.
96.4%
Steep 4B Balanced picks the same next token as the full-precision original, at 3.2 GB on disk instead of 8.7 GB.
87.8%
Crema 6B Balanced, our own int8 conversion: by eye, its pictures cannot be told from the original’s. 6.2 GB instead of 12.3 GB.
12
builds in the catalogue, each listed with what it was measured against, and what was not measured yet.
wikitext-2 test set, 32 × 512 tokens, against the full-precision original. The importance matrix halves the error of q4: KL 0.0374 without it, 0.0202 with it.
6 prompts at 1024 × 1024, 8 steps, fixed seed, against the full-precision original. The smaller two lose fine texture; the smallest also shifts the composition, though its pictures are still usable.
Whistle 82M: smaller was worse
We tried shrinking it. On the processor, the smaller builds were slower and made more mistakes, so it ships at full precision.
| Build | Size | Word errors | Time per second of speech |
|---|---|---|---|
| Full precision (ships) | 326 MB | 6.2% | 0.75 s |
| int8, matrix layers | 279 MB | 6.9% | 0.97 s |
| int8, everything | 114 MB | 8.5% | 2.96 s |
10 sentences, one voice, on the processor; word error rate after transcribing the audio back to text, one processor thread.
Pour 5B: not measured yet
Both builds ran and the motion looked coherent. Quality against the original has not been measured yet. Until it is, the catalogue says n/m instead of a number.
Test: 33 frames at 480 × 832, 20 steps.
About the speeds
The lab machine was busy with other builds during these runs, so lab timings only compare builds with each other. In the chat above, on the same MacBook Pro, Steep ran at 113.3 tokens per second.
Every build, every number
| Build | Version | Memory | Download | Quality | Measured as |
|---|---|---|---|---|---|
| Steep 4B q3 | Smallest | 3.3 GB | 2.5 GB | 86.9% | same next token; KL 0.0819; 62.5 tok/s |
| Steep 4B q4 | Small | 3.7 GB | 3 GB | 93.2% | same next token; KL 0.0202; 66.3 tok/s |
| Steep 4B q5 | Balanced | 4.1 GB | 3.3 GB | 96.4% | same next token; KL 0.0066; 54.6 tok/s |
| Steep 4B q6 | Better | 4.4 GB | 3.7 GB | 97.3% | same next token; KL 0.0026; 50.4 tok/s |
| Steep 4B q8 | Best quality | 5.4 GB | 4.6 GB | 99.1% | same next token; KL 0.0005; 52.4 tok/s |
| Crema 6B q4 | Smallest | 7.2 GB | 5.7 GB | 66.7% | 1 − LPIPS; SSIM 0.710; PSNR 16.3 dB |
| Crema 6B q5 | Small | 7.9 GB | 6.4 GB | 77.0% | 1 − LPIPS; SSIM 0.770; PSNR 18.9 dB |
| Crema 6B int8 | Balanced | 9.7 GB | 8.3 GB | 87.8% | 1 − LPIPS; SSIM 0.858; PSNR 22.8 dB |
| Crema 6B q8 | Best quality | 10 GB | 8.6 GB | 88.3% | 1 − LPIPS; SSIM 0.861; PSNR 22.8 dB |
| Pour 5B q4 | Smaller | 11.3 GB | 8.3 GB | n/m | not measured: 33 frames, 480x832, 20 steps: ran and looked coherent |
| Pour 5B q8 | Better quality | 16.2 GB | 13.3 GB | n/m | not measured: 33 frames, 480x832, 20 steps: ran and looked coherent |
| Whistle 82M f32 | Standard | 700 MB | 384 MB | 100.0% | weights unchanged; word errors 6.2% |
From the package manifest; how each kind of model is measured is in docs/models.md. Memory figures are estimates.
Each of these runs costs hours on a fast machine. Sponsors pay for the measurements, and every result is published, including the ones that did not work out.
04 Families
Five families. One name each.
Rigbrew converts and measures open-weight models in its own lab, then ships them under its own names. Only models under Apache-2.0, MIT or CC-BY-4.0 are packaged, and every package carries its licence and notice files.
- 01
Steep
Chat
Answers questions, writes and explains. The model you talk to. It can call the other installed families for you.
Steep 4B ·5 builds ·3.3–5.4 GB ·Apache-2.0
Runs today - 02
Crema
Pictures
Makes a picture from a description. Adds Picture mode to the chat, and Steep can ask it for a picture when you want one.
Crema 6B ·4 builds ·7.2–10 GB ·Apache-2.0
Runs today - 03
Pour
Video
Makes short video clips from a description. Downloads today. Running it inside Rigbrew comes later.
Pour 5B ·2 builds ·11.3–16.2 GB ·Apache-2.0
Download now, runs later - 04
Whistle
Voice
Reads text aloud. Downloads today. Reading answers aloud in the chat comes later.
Whistle 82M ·1 build ·700 MB ·Apache-2.0
Download now, runs later - 05
Sift
Listening
Turns speech into text. Planned: dictate instead of typing.
No package yet
Planned
Modules are drawn to scale: the memory of the version Rigbrew picks on a 48 GB Mac, on a ruler that ends at 18 GB.
05 Start simple, grow pro
Plain words first. Every number when you want it.
A new install starts at Guided: names by what they do, one button per model. One setting shows more, up to every build with its measurements. Below, the same Models screen at each level.
- Names by what they do: “Chat (Steep 4B)”
- Only the version Rigbrew picked for this machine
- “Fast on this Mac” instead of tokens per second
- A tour, a checklist and tips
One setting changes every screen. The app may suggest the next level once you use it more; it never switches by itself, and switching installs or removes nothing.






06 Privacy and open source
Runs here. Says what it sends.
The models run on your computer and keep working offline. The app does report to Rigbrew’s own server, in two layers, and lists everything it has queued to send.
What you type is never sent to a model provider.
- Always Basic usage
- What was opened, installed and loaded, how fast it ran, errors, the hardware class, the app version and a random install ID. Never what you type or what the models answer. Not linked to an account, and the IP address is not stored.
- With consent Content
- The text of prompts and answers, linked to an account so it can be exported or deleted. The app asks for this at first start and does not run without it; withdraw it in Settings and whatever has not been sent is deleted. Builds from source have no account module and never send content.
- Pictures, audio and video you make are never uploaded.
- Settings › Privacy and data lists every event before it is sent.
- Developers switch basic usage off with
DO_NOT_TRACK=1. - This site counts page views without cookies and stores nothing on your device.
Read the Privacy Policy (a draft, not yet reviewed by a lawyer)
Apache-2.0. Read every line that runs.
rigbrew/ ├─ core/ Rust: scan, selector, downloader, engines ├─ daemon/ rigbrewd, the local service ├─ cli/ rigbrew on the command line ├─ app/ the desktop app, Tauri 2 and SvelteKit └─ docs/ architecture, the local API, models
The ruler at the top of this page runs the same budget rules as tier.rs, and the code that sends usage data is in events.rs and cloud.rs.
07 In the open
Built in the open. Counted in the open.
These numbers come live from Rigbrew’s server, as they are, starting from zero. Sponsors pay for the machine hours behind every build and every measurement.
Live from Rigbrew’s server
- GitHub stars
- –
- Contributors
- –
- Installs
- –
- Downloads
- –
- Sponsors
- –
Counted from the first day, not rounded up. A zero shows as new.
Sponsors
Rigbrew is free and open. Sponsors pay for the builds and the measurements.
No sponsors yet. The first name here could be yours.
08 Get started
Not downloadable yet. Getting close.
macOS on Apple Silicon comes first. Until the beta is signed and published, Rigbrew runs from source, and the steps are short.
- macOS Apple Silicon Beta coming Runs from source today. A signed download is next.
- Windows Graphics card or not Next After macOS. The budget rules above already cover graphics cards.
- Linux Later Planned after Windows.
To hear when the beta is out, watch the repository’s releases on GitHub. There is no mailing list.