diff --git a/.gitignore b/.gitignore
index 04c7330..b552e16 100644
--- a/.gitignore
+++ b/.gitignore
@@ -4,3 +4,4 @@ claude.md
 **.pyc
 __pycache__/*
 *.egg-info/*
+/models
diff --git a/AGENTS.md b/AGENTS.md
new file mode 100644
index 0000000..b60b198
--- /dev/null
+++ b/AGENTS.md
@@ -0,0 +1,33 @@
+# Repository Guidelines
+
+## Project Structure & Module Organization
+- `mlx_video/` holds the core package. Entry points live in `mlx_video/generate.py` and `mlx_video/generate_av.py`, with utilities in `mlx_video/utils.py` and `mlx_video/postprocess.py`.
+- Model code is under `mlx_video/models/ltx/`, including VAE components in `video_vae/` and `audio_vae/`, plus prompt templates in `prompts/`.
+- Tests are in `tests/` (e.g., `tests/test_rope.py`, `tests/test_vae_streaming.py`).
+- Sample output media lives in `examples/`.
+
+## Build, Test, and Development Commands
+- Install (editable): `pip install -e '.[dev]'` or `uv pip install -e '.[dev]'` (installs pytest).
+- Run the CLI: `uv run mlx_video.generate --prompt "..." -n 100 --width 768`.
+- Module execution: `python -m mlx_video.generate --prompt "..." --height 768 --width 768`.
+- Tests: `pytest` from the repo root.
+
+## Coding Style & Naming Conventions
+- Python code uses 4-space indentation and standard PEP 8 conventions.
+- Prefer descriptive, lowercase module and function names (`generate_av.py`, `decode_with_tiling`).
+- Tests follow `test_*.py` naming and `Test*` classes with `test_*` methods.
+- No formatter/linter is configured in-repo; keep formatting consistent with existing files.
+
+## Testing Guidelines
+- Test framework: pytest (declared in `pyproject.toml` under `dev` extras).
+- Keep tests deterministic by seeding MLX where applicable (see `mx.random.seed(42)` patterns).
+- Cover shape invariants and numerical stability; use `np.testing.assert_allclose` for tensor checks.
+
+## Commit & Pull Request Guidelines
+- Recent history uses short, imperative summaries (e.g., “add tests”, “Enhance video generation…”). Keep the subject concise and action-oriented.
+- PRs should include a brief description of changes, testing performed (commands + results), and any relevant sample outputs (e.g., generated `.mp4` or `.gif`).
+- If your change alters model outputs or performance, note the expected impact and any new flags.
+
+## Runtime & Environment Notes
+- Target environment is macOS on Apple Silicon with Python >= 3.11 and MLX >= 0.22.0.
+- Model weights are expected from Hugging Face (default `Lightricks/LTX-2`); ensure access and disk space before runs.
diff --git a/README.md b/README.md
index 190bdf7..9031789 100644
--- a/README.md
+++ b/README.md
@@ -16,6 +16,20 @@ pip install git+https://github.com/Blaizzy/mlx-video.git
 uv pip install git+https://github.com/Blaizzy/mlx-video.git
 ```
 
+## Model Downloads
+
+Use the repo script to prefetch weights into the Hugging Face cache (so `snapshot_download` can find them locally):
+
+```bash
+./download.sh
+```
+
+You can also set a custom cache location:
+
+```bash
+./download.sh --hf-home /path/to/hf-cache
+```
+
 Supported models:
 
 ### LTX-2
diff --git a/download.sh b/download.sh
new file mode 100755
index 0000000..fc95335
--- /dev/null
+++ b/download.sh
@@ -0,0 +1,137 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+MODEL_REPO="Lightricks/LTX-2"
+TEXT_ENCODER_REPO="google/gemma-3-12b-it"
+SKIP_TEXT_ENCODER=0
+HF_HOME_DIR=""
+HF_MIRROR="https://hf-mirror.com"
+
+usage() {
+  cat <<'USAGE'
+Usage: ./download.sh [options]
+
+Options:
+  --model-repo REPO          Model repo to download (default: Lightricks/LTX-2)
+  --text-encoder-repo REPO   Text encoder repo to download (default: google/gemma-3-12b-it)
+  --skip-text-encoder        Skip downloading the text encoder repo
+  --hf-home DIR              Set HF_HOME for Hugging Face cache location
+  -h, --help                 Show this help
+
+Notes:
+  - Set HF_TOKEN in your environment if the repo is gated or requires auth.
+  - The script uses HF mirror by default and automatically falls back to
+    https://huggingface.co when the mirror returns HTTP 403.
+  - Downloads go to the Hugging Face cache so the repo can find them.
+USAGE
+}
+
+while [[ $# -gt 0 ]]; do
+  case "$1" in
+    --model-repo)
+      MODEL_REPO="$2"
+      shift 2
+      ;;
+    --text-encoder-repo)
+      TEXT_ENCODER_REPO="$2"
+      shift 2
+      ;;
+    --skip-text-encoder)
+      SKIP_TEXT_ENCODER=1
+      shift
+      ;;
+    --hf-home)
+      HF_HOME_DIR="$2"
+      shift 2
+      ;;
+    -h|--help)
+      usage
+      exit 0
+      ;;
+    *)
+      echo "Unknown option: $1" >&2
+      usage
+      exit 1
+      ;;
+  esac
+done
+
+if command -v python3 >/dev/null 2>&1; then
+  PYTHON=python3
+else
+  PYTHON=python
+fi
+
+export HF_ENDPOINT="$HF_MIRROR"
+if [[ -n "$HF_HOME_DIR" ]]; then
+  export HF_HOME="$HF_HOME_DIR"
+fi
+
+download_repo() {
+  local repo="$1"
+  shift 1
+  local patterns_str
+  patterns_str=$(IFS='|'; echo "$*")
+
+  REPO="$repo" ALLOW_PATTERNS="$patterns_str" "$PYTHON" - <<'PY'
+import os
+from huggingface_hub import snapshot_download
+
+repo = os.environ["REPO"]
+patterns = os.environ.get("ALLOW_PATTERNS")
+allow_patterns = patterns.split("|") if patterns else None
+mirror = os.environ.get("HF_ENDPOINT")
+
+def run_download(endpoint):
+    if endpoint:
+        os.environ["HF_ENDPOINT"] = endpoint
+        label = endpoint
+    else:
+        os.environ.pop("HF_ENDPOINT", None)
+        label = "https://huggingface.co"
+    print(f"==> Downloading {repo} via {label}")
+    return snapshot_download(
+        repo_id=repo,
+        resume_download=True,
+        allow_patterns=allow_patterns,
+    )
+
+try:
+    run_download(mirror)
+except Exception as err:  # pragma: no cover - shell script helper
+    status = getattr(getattr(err, "response", None), "status_code", None)
+    is_403 = status == 403 or "403" in str(err)
+    if is_403:
+        print("Mirror returned 403; retrying with https://huggingface.co")
+        run_download("https://huggingface.co")
+    else:
+        raise
+PY
+}
+
+download_repo "$MODEL_REPO" \
+  "*.safetensors" \
+  "*.json" \
+  "tokenizer/**" \
+  "tokenizer.*" \
+  "special_tokens_map.json" \
+  "tokenizer_config.json" \
+  "tokenizer.json" \
+  "vocab.json" \
+  "merges.txt"
+
+if [[ "$SKIP_TEXT_ENCODER" -eq 0 ]]; then
+  download_repo "$TEXT_ENCODER_REPO" \
+    "*.safetensors" \
+    "*.json" \
+    "tokenizer/**" \
+    "tokenizer.*" \
+    "special_tokens_map.json" \
+    "tokenizer_config.json" \
+    "tokenizer.json" \
+    "tokenizer.model" \
+    "vocab.json" \
+    "merges.txt"
+fi
+
+echo "Done. Downloaded models are in the Hugging Face cache."