Get OpenUnum. On your machine.
Which model and features OpenUnum gives you on your hardware, the three machines we tested it on, the full setup of our own box with its traps, and the install command.
What runs on my machine?
Pick the kind of machine and its memory. The answer is the same table the public installer uses when it reads your hardware, so the installer will tell you the same thing.
| Memory (RAM) | Recommended model | With a GPU | Context (tokens) | Note |
|---|
| from 4 GB to 5 GB | Qwen 2.5 3B (about 2.0 GB) | Qwen 2.5 3B (about 2.0 GB) | 4,096 | Very tight. Expect it to be slow; a hosted model you choose is the better route on this machine. |
| over 5 GB to 11 GB | Qwen 2.5 3B (about 2.0 GB) | Qwen 2.5 3B (about 2.0 GB) | 8,192 | A small model for short tasks. The control model fits beside it from 6 GB. |
| over 11 GB to 22 GB | Qwen 2.5 3B (about 2.0 GB) | SuperGemma 4 (about 5.4 GB) | 8,192 | With 6 GB of VRAM or more the installer moves up to the larger model on the GPU. |
| over 22 GB to 47 GB | SuperGemma 4 (about 5.4 GB) | SuperGemma 4 (about 5.4 GB) | 8,192 (12,288) | Comfortable for everyday agent work. |
| over 47 GB to 96 GB | Qwen 3 8B (about 4.8 GB) | Qwen 3 8B (about 4.8 GB) | 16,384 (32,768 with 12 GB of VRAM or more) | Room for longer contexts and a second model beside the main one. |
| over 96 GB | Qwen 3 8B (about 4.8 GB) by default | Qwen 3 8B (about 4.8 GB) by default | 32,768 | The default download stays a safe 8B, but this class can run much larger mixture-of-experts models from the in-app model catalogue. That is what our own box does. |
- Agent runtime, tools, memory, verification and recovery: yes. The runtime is light; the model is the limit.
- OpenUnum control model (1.7B, routes and checks work beside the chat model): yes, from 6 GB of RAM.
- OpenUnum control model: not on this machine (needs 6 GB of RAM).
- Hosted models: optional on every tier, off unless you add a key.
- Image, video, music and voice generation: yes. This is the class we run it on.
- Image generation: yes, slowly (about 2.5 minutes a picture on our handheld). Video, music and voice: not tested here.
- Image, video, music and voice generation: not tested on this class.
- Under 4 GB there is no local model. The runtime still runs with a hosted model you choose.
- macOS installs from source only, and the source is private (access on request).
- On a phone the small model is for offline use; real work goes to a bigger model on your own box or a hosted route you choose.
Source: the tier table in the public install script (last reviewed 2026-06-19). Model sizes are download sizes. Speed depends on the machine and is published only for hardware we measured.
Tested, testing, should work
"Tested" means we run OpenUnum on that exact machine. "Testing" means a run is in progress and there are no numbers yet. "Should work" means the installer supports the platform and nobody here has checked it.
- Bosgame M5 mini PC (Tested by us), AMD Ryzen AI Max+ 395 (Strix Halo), 128 GB unified memory, Ubuntu 24.04: The whole lab on one box: a large local model at 131k context for two agents at once, the always-on agent crew, file search, and image, video, music and voice generation. One heavy GPU job at a time. Decode 43.2 tokens/s median in production on 2026-11-10. Bosgame M5 product page
- ROG Ally X (Tested by us), Handheld gaming PC, Linux: OpenUnum every day with a small local model, kept current through the same public release channel as any install. It also generates images (about 2.5 minutes for a 576 x 1024 picture) and takes media work from the big box when that one is busy. No model-speed numbers published yet.
- Samsung Galaxy S22 (Tested by us), Galaxy S22 booting our own Linux (unumS22linux, open source): The agent, push-to-talk voice (about 7 seconds from button to spoken answer) and the phone's hardware as tools. A small model runs on the phone; for real work it uses a bigger model somewhere else: your own box, or a hosted route you choose. Camera capture is not working yet. unumS22linux project
- Windows laptop, 16 GB, Intel integrated graphics (In testing, no numbers yet), Windows 11, 16 GB RAM, Intel iGPU: In testing now with the public Windows installer. No numbers yet.
- Mac (Apple Silicon) (Should work, not tested by us): Installs from source only today, and the source is private: access on request. No prebuilt macOS bundle yet.
- Gaming PC with a discrete GPU (Should work, not tested by us): The installer reads your VRAM and offloads the model to the GPU from 4 GB of VRAM up. Linux and Windows builds exist.
- NVIDIA DGX Spark (Should work, not tested by us): A Linux arm64 build exists and the machine is in the same big-unified-memory class as our box. Nobody here has run it.
- Other 16 GB laptops (Should work, not tested by us): The installer picks a 3B model on the CPU, or a larger one when it finds 6 GB of VRAM or more.
Install
Linux (x64, arm64)
bash <(curl -fsSL https://install.corpunum.com/install.sh)
Windows (PowerShell)
powershell -ExecutionPolicy Bypass -Command "iwr -useb https://install.corpunum.com/install-windows.ps1 | iex"
View install.sh · View install-windows.ps1 · Update, uninstall, macOS and every option
The full setup of our box
This is what runs in the lab today, to copy or to compare with yours.
One mini PC
- Bosgame M5 mini PC: AMD Ryzen AI Max+ 395 (Strix Halo) with the Radeon 8060S graphics on the same chip.
- 128 GB of unified memory, shared between the processor and the graphics. 2 TB NVMe, plus a second disk for models.
- Ubuntu 24.04. No discrete GPU, no cloud inference.
One large local model
- A Qwen-family mixture-of-experts model (the Qwen 3.8 Flash Next line, a community variant) at 4-bit quantisation, with its multi-token-prediction head and a vision projector.
- About 68 GiB of weights in graphics memory.
- Context 131,072 tokens per request, two slots, so two agents work at once. Default temperature 0.
Settings to copy
- On an open-source engine: context 131,072, 2 parallel slots, weights locked in memory with an unlimited memlock limit, an 8-bit KV cache, a prompt cache of 36 entries, temperature 0.
- What we run in production since 2026-11-10 is our own engine. It is private for now and runs only on this chip family. The numbers, wins and losses, are public.
- Total graphics memory in use: about 75 GiB with the open-source engine, about 81 GiB with ours.
The numbers
What bit us
- Lock the weights. Without it the kernel takes back the pages the graphics side has mapped, and the model stalls for minutes. With the lock we measured 1.5 seconds of evictions in 11 minutes; before it, 820 seconds in 18 minutes.
- Set vm.compact_unevictable_allowed=0, and check after every reboot that the locked amount is what you expect (about 74 GB here).
- Drop caches and compact memory before you load the model. A load into fragmented memory swaps and then decodes at 0.35 to 12 tokens/s.
- Do not raise vm.min_free_kbytes or vm.watermark_scale_factor: the weights then fail to pin.
- One heavy job at a time. A compiler run or a video render beside the model causes the same evictions. We use one lock file that every heavy job must hold.
What the agent works with
- OpenUnum itself: the agent runtime, from the public installer below.
- unumsearch for file search, OpenLunum for context, UnumPingPong for worker and referee loops. All three are open source.
Skills and tools, with success rates