Models
The model sheet lists what can run on this device, with speed, download size and context on every row. Two kinds of model appear: OIOXO models, which we build and which count against the free window, and open models, which are stock releases from Hugging Face and never count against anything.
OIOXO models
| Model | Download | Needs free memory | Notes |
|---|---|---|---|
| OIOXO Light | 540 MB | 750 MB | Fastest; the default on most laptops. |
| OIOXO Prime | 1.16 GB | 1.5 GB | The stronger OIOXO model. |
| OIOXO Ultra | 2.5 GB | 2.8 GB | Desktops and strong GPUs. |
| OIOXO Light (Mobile) | 545 MB | 750 MB | The phone build of Light. |
On Free, OIOXO models are available for a 5-hour window with weekly and monthly caps; on Pro they are unlimited. The meter is always visible before a limit is hit.
Open models
The catalog runs from 360M to 32B. Every model is offered on every machine: the picker measures your device and marks a row that will not fit, with the reason — so a 32B on a laptop is a visible, explained row rather than a hidden one, and the same account on a bigger machine simply runs it.
Runs in a browser tab
| Model | Download | Context | Notes |
|---|---|---|---|
| SmolLM2 360M | 390 MB (Q8_0) | 8k | Chat only. Runs on almost anything. |
| Qwen2.5 Coder 0.5B | 400 MB (Q4_K_M) | 32k | Chat and completion. |
| Llama 3.2 1B | 810 MB (Q4_K_M) | 128k | Chat. |
| Qwen2.5 Coder 1.5B | 1.1 GB (Q4_K_M) | 32k | Agent-capable; wants 4 GB RAM. |
| Qwen2.5 Coder 3B | 1.8 GB (Q4_K_M) | 32k | Agent-capable; wants 8 GB RAM. |
Desktop app
The native engine has no browser memory ceiling, so the rest of the range opens up. All Q4_K_M, all agent-capable.
| Model | Download | Context | Notes |
|---|---|---|---|
| Phi-4 Mini 3.8B | 2.4 GB | 128k | Microsoft. |
| DeepSeek Coder 6.7B | 3.9 GB | 16k | Coding. |
| Qwen2.5 Coder 7B | 4.5 GB | 32k | The strongest coder that still fits a mid-range machine. |
| Qwen3 8B | 4.8 GB | 40k | General and coding. |
| Gemma 3 12B | 7.0 GB | 128k | Google. |
| Mistral Nemo 12B | 7.1 GB | 128k | Long context. |
| Qwen2.5 Coder 14B | 8.6 GB | 32k | Coding. |
| Qwen3 14B | 8.6 GB | 40k | General and coding. |
| Phi-4 14B | 8.6 GB | 16k | Microsoft. |
| Devstral 24B | 13.7 GB | 128k | Mistral, built for agentic coding. |
| Gemma 3 27B | 15.8 GB | 128k | Google. |
| Qwen3 32B | 18.8 GB | 40k | Needs a workstation. |
| Qwen2.5 Coder 32B | 18.9 GB | 32k | The top open coder. Needs a workstation. |
| OIOXO Ultra 35B | 21.1 GB | 256k | Sparse MoE architecture (3B active parameters). Full quality workstation reasoning. |
Open models are unlimited on every plan. The web sheet hides anything over 2.2 GB or 9B parameters, because a browser tab cannot hold it — those rows are what the desktop app adds.
Helpers that download only when used
OIOXO Vision (260 MB) and Vision Prime (511 MB) read an image you attach to the composer. They are fetched the first time a feature needs them, never on boot.
Consent and fit
- Nothing downloads without a click. A model that does not fit is shown as such and not started.
- A downloaded model stays in browser storage and loads from disk on the next visit. Settings → Models & keys lists and deletes them.
- A first run can take minutes on CPU while the model reads the prompt; the strip shows the estimate and, past it, the elapsed time.
Start with what the sheet recommends. If a run writes a file that does not work, OIOXO offers a bigger model you already have or one of your keys — see Modes.