diff --git a/backend/test_cuda_compression.py b/backend/test_cuda_compression.py
new file mode 100644
index 00000000..079f8be1
--- /dev/null
+++ b/backend/test_cuda_compression.py
@@ -0,0 +1,137 @@
+#!/usr/bin/env python3
+"""
+Test CUDA binary compression to verify it fits under GitHub's 2GB release asset limit.
+
+Usage:
+    python test_cuda_compression.py [path/to/voicebox-server-cuda.exe]
+
+If no path provided, looks for the binary in ./dist/
+"""
+
+import os
+import sys
+import subprocess
+from pathlib import Path
+
+
+def format_size(bytes_size):
+    """Format bytes into human-readable size."""
+    for unit in ['B', 'KB', 'MB', 'GB']:
+        if bytes_size < 1024.0:
+            return f"{bytes_size:.2f} {unit}"
+        bytes_size /= 1024.0
+    return f"{bytes_size:.2f} TB"
+
+
+def get_file_size(filepath):
+    """Get file size in bytes."""
+    return os.path.getsize(filepath)
+
+
+def compress_with_7z(input_file, output_file):
+    """Compress file using 7z with maximum compression."""
+    print(f"\nCompressing with 7z (maximum compression)...")
+    print(f"This may take several minutes for a ~2.5GB file...\n")
+
+    cmd = [
+        '7z', 'a',
+        '-t7z',           # 7z format
+        '-m0=lzma2',      # LZMA2 compression
+        '-mx=9',          # Maximum compression
+        '-mfb=64',        # Fast bytes
+        '-md=32m',        # Dictionary size
+        '-ms=on',         # Solid archive
+        output_file,
+        input_file
+    ]
+
+    try:
+        subprocess.run(cmd, check=True, capture_output=True, text=True)
+        return True
+    except subprocess.CalledProcessError as e:
+        print(f"Error during compression: {e}")
+        print(f"stderr: {e.stderr}")
+        return False
+    except FileNotFoundError:
+        print("ERROR: 7z not found. Please install 7-Zip:")
+        print("  Windows: https://www.7-zip.org/download.html")
+        print("  macOS: brew install p7zip")
+        print("  Linux: apt-get install p7zip-full")
+        return False
+
+
+def main():
+    # Find CUDA binary
+    if len(sys.argv) > 1:
+        cuda_binary = Path(sys.argv[1])
+    else:
+        # Look in dist directory
+        dist_dir = Path(__file__).parent / 'dist'
+        candidates = list(dist_dir.glob('voicebox-server-cuda*.exe'))
+
+        if not candidates:
+            print("ERROR: CUDA binary not found in ./dist/")
+            print("Please provide the path as an argument:")
+            print("  python test_cuda_compression.py path/to/voicebox-server-cuda.exe")
+            sys.exit(1)
+
+        cuda_binary = candidates[0]
+
+    if not cuda_binary.exists():
+        print(f"ERROR: File not found: {cuda_binary}")
+        sys.exit(1)
+
+    print("=" * 70)
+    print("CUDA Binary Compression Test")
+    print("=" * 70)
+
+    # Get original size
+    original_size = get_file_size(cuda_binary)
+    print(f"\nOriginal file: {cuda_binary.name}")
+    print(f"Original size: {format_size(original_size)} ({original_size:,} bytes)")
+
+    # Check if already over 2GB
+    github_limit = 2 * 1024 * 1024 * 1024  # 2GB in bytes
+    print(f"GitHub limit:  {format_size(github_limit)} ({github_limit:,} bytes)")
+
+    if original_size > github_limit:
+        print(f"\n⚠️  Original file exceeds GitHub limit by {format_size(original_size - github_limit)}")
+    else:
+        print(f"\n✓ Original file is under GitHub limit")
+
+    # Compress
+    output_file = cuda_binary.parent / f"{cuda_binary.stem}.7z"
+    if output_file.exists():
+        print(f"\nRemoving existing compressed file: {output_file.name}")
+        output_file.unlink()
+
+    success = compress_with_7z(cuda_binary, output_file)
+
+    if not success:
+        sys.exit(1)
+
+    # Check compressed size
+    compressed_size = get_file_size(output_file)
+    compression_ratio = (1 - compressed_size / original_size) * 100
+
+    print("\n" + "=" * 70)
+    print("Compression Results")
+    print("=" * 70)
+    print(f"\nCompressed file: {output_file.name}")
+    print(f"Compressed size: {format_size(compressed_size)} ({compressed_size:,} bytes)")
+    print(f"Compression ratio: {compression_ratio:.1f}%")
+    print(f"Space saved: {format_size(original_size - compressed_size)}")
+
+    if compressed_size <= github_limit:
+        print(f"\n✅ SUCCESS: Compressed file fits under GitHub's 2GB limit!")
+        print(f"   Margin: {format_size(github_limit - compressed_size)} remaining")
+    else:
+        print(f"\n❌ FAILED: Compressed file still exceeds GitHub limit")
+        print(f"   Over by: {format_size(compressed_size - github_limit)}")
+        print(f"\n   Alternative: Host on external storage (S3, Azure Blob, etc.)")
+
+    print("\n" + "=" * 70)
+
+
+if __name__ == '__main__':
+    main()
diff --git a/docs/github-2gb-limit-issue.md b/docs/github-2gb-limit-issue.md
new file mode 100644
index 00000000..0308de40
--- /dev/null
+++ b/docs/github-2gb-limit-issue.md
@@ -0,0 +1,122 @@
+# GitHub 2GB Release Asset Limit Issue
+
+## Problem
+
+The CUDA server binary upload fails in CI with:
+```
+Error: File size (2543828017) is greater than 2 GiB
+```
+
+GitHub release assets have a hard limit of 2GB per file. Our CUDA binary is ~2.5GB, which exceeds this limit.
+
+## Background
+
+The dual-server binary system (see `dual-server-binaries.md`) creates two binaries:
+- **CPU binary**: ~500MB ✅ Works fine
+- **CUDA binary**: ~2.5GB ❌ Exceeds GitHub limit
+
+## Attempted Solution: Compression
+
+We're testing 7z compression with maximum settings to see if we can squeeze the CUDA binary under 2GB.
+
+### Test Script
+
+Run `backend/test_cuda_compression.py` to test compression locally:
+
+```bash
+cd backend
+python test_cuda_compression.py
+```
+
+This will:
+1. Find the CUDA binary in `dist/`
+2. Compress it with 7z (maximum compression)
+3. Report if the compressed size fits under 2GB
+
+### Expected Compression
+
+PyTorch CUDA binaries typically compress well since they contain:
+- Repeated patterns in neural network weights
+- Debug symbols and metadata
+- Redundant CUDA libraries
+
+Estimated compression: 30-40% reduction
+- Original: ~2.5GB
+- Target: <2GB
+- Required compression: >20%
+
+## Fallback: External Hosting
+
+If compression doesn't work, we'll need to host the CUDA binary externally:
+
+### Option 1: AWS S3
+```yaml
+- name: Upload CUDA binary to S3
+  run: |
+    aws s3 cp backend/cuda-release/voicebox-server-cuda-*.exe \
+      s3://voicebox-releases/cuda-binaries/${{ github.ref_name }}/
+```
+
+### Option 2: Azure Blob Storage
+```yaml
+- name: Upload to Azure Blob
+  run: |
+    az storage blob upload \
+      --account-name voiceboxreleases \
+      --container-name cuda-binaries \
+      --file backend/cuda-release/voicebox-server-cuda-*.exe
+```
+
+### Option 3: GitHub Packages (Container Registry)
+Package as a container image, though this adds complexity for desktop app distribution.
+
+## Implementation Plan
+
+1. **Test compression locally** ← Current step
+2. **If compression works (<2GB)**:
+   - Update CI to compress before upload
+   - Update app to handle .7z downloads
+   - Add extraction step in download manager
+
+3. **If compression fails (≥2GB)**:
+   - Set up external storage (likely S3)
+   - Update CI to upload to S3
+   - Provide download URL in release notes
+   - Update app download manager to fetch from S3
+
+## CI Workflow Changes (if compression works)
+
+```yaml
+- name: Compress CUDA binary (Windows only)
+  if: matrix.platform == 'windows-latest'
+  shell: bash
+  run: |
+    cd backend/cuda-release
+    7z a -t7z -m0=lzma2 -mx=9 -mfb=64 -md=32m -ms=on \
+      voicebox-server-cuda-x86_64-pc-windows-msvc.7z \
+      voicebox-server-cuda-*.exe
+
+- name: Upload compressed CUDA server (Windows only)
+  if: matrix.platform == 'windows-latest'
+  uses: softprops/action-gh-release@v1
+  with:
+    files: backend/cuda-release/*.7z
+```
+
+## User Experience Impact
+
+### With Compression
+- Download: `voicebox-server-cuda-*.7z` (~1.5-1.8GB)
+- App extracts automatically
+- One extra step but manageable
+
+### With External Hosting
+- Download from S3/Azure URL
+- No GitHub release asset dependency
+- Potentially faster download speeds (CDN)
+
+## Status
+
+🔄 **Testing compression locally to determine viability**
+
+Results pending from local test run.