# Context size

ALIA-40b has a **native context of 163840 tokens (~160K)** (`rope_theta 1e4`).

Included Modelfiles set `num_ctx`:
- `Modelfile.8k`   — 8192 (lightest KV RAM)
- `Modelfile.32k`  — 32768 (good default)
- `Modelfile.128k` — 131072 (long documents; needs a lot of RAM for KV)

```bash
ollama create alia-40b-32k -f Modelfile.32k && ollama run alia-40b-32k
```

A 46B at long context needs substantial RAM: weights + KV-cache (KV grows with context). At 128K the KV alone is large — budget accordingly, or use a smaller quant to leave room.