pocketai-cdn

Static CDN serving model catalogs for PocketAI apps. Hosted via GitHub Pages at cdn.pocketaihub.com.

CDN vs GitHub Releases

The CDN remains the live source of model catalogs and catalog version metadata. GitHub Releases hosts large PocketAI app and runtime binaries, while model weights continue to download from their pinned upstream repositories (primarily Hugging Face). Moving binary releases to GitHub did not replace the catalog CDN.

Base URL

https://cdn.pocketaihub.com

Endpoints

Language Models

GET https://cdn.pocketaihub.com/catalog.json

Returns the full language model catalog. Each entry includes model ID, name, provider, size, RAM/VRAM requirements, download URL, supported platforms, capabilities (e.g. vision), and context length.

Image Models

GET https://cdn.pocketaihub.com/image-catalog.json

Returns the image generation model catalog (Stable Diffusion, FLUX, SDXL, Chroma, Z-Image, Ovis-Image, Qwen-Image). Each entry includes architecture, supported resolutions, default steps/CFG, and auxiliary model URLs (CLIP, T5, VAE).

MLX Image Models

GET https://cdn.pocketaihub.com/mlx-image-catalog.json

Returns Apple Silicon image generation models that load through MLX-native runtimes instead of the GGUF/stable-diffusion.cpp path. Each entry includes the Hugging Face repo id, snapshot revision, backend/runtime selector, quantization, supported resolutions, and default generation settings.

Video Models

GET https://cdn.pocketaihub.com/video-catalog.json

Returns the video generation model catalog (including Wan, LTX, Kandinsky, HunyuanVideo, and MiniMax H3 families). Each entry includes mode (t2v, i2v, flf2v), supported resolutions, frame counts, FPS, and auxiliary model URLs (text encoders, video/audio VAEs, and CLIP Vision).

Voice Models

GET https://cdn.pocketaihub.com/voice-catalog.json

Returns the voice model catalog covering:

• STT - Whisper models (Tiny through Large v3 Turbo)
• TTS - Kokoro and Piper text-to-speech models
• VAD - Silero voice activity detection

Audio Generation Models

GET https://cdn.pocketaihub.com/audio-catalog.json

Returns music and sound-effect generation models. Entries pin their upstream repository revisions and SHA-256-verified artifacts. Model weights download directly from the publisher; this CDN hosts metadata only. Discovery-only entries intentionally omit install artifacts until their runtime admission is complete.

Gated Models

GET https://cdn.pocketaihub.com/gated-catalog.json

Returns gated/uncensored models separated from the main catalogs. Structured as { "image": [...], "video": [...], "language": [...] }. Only fetched by non-app-store builds — app store builds never request this endpoint.

Catalog Versions

GET https://cdn.pocketaihub.com/catalog-versions.json

Returns the current version number for each catalog category. Used by apps to check for updates without downloading the full catalogs.

{
  "language": 1,
  "image": 1,
  "video": 1,
  "voice": 1,
  "audio": 1,
  "gated": 1,
  "mesh": 1,
  "mlx": 1,
  "mlxImage": 1
}

Model Weight Tiers

Each language model entry includes a weight field indicating its download size class:

Weight	Size Range	Typical Use
tiny	< 1 GB	Older phones, fast downloads
light	1 - 3 GB	Most phones and tablets
medium	3 - 10 GB	Flagship phones, all desktops
large	10 - 25 GB	Desktop / laptop only
heavy	25 - 80 GB	High-memory workstations
massive	80 GB+	Multi-GPU / high-end setups

Where Catalogs Are Stored

There are two copies of the model catalogs that must stay in sync:

Location	Description
pocketai-cdn/catalog.json	CDN (source of truth) - language model catalog served at cdn.pocketaihub.com/catalog.json
pocketai-cdn/image-catalog.json	CDN - image generation model catalog
pocketai-cdn/mlx-image-catalog.json	CDN - MLX-native image generation model catalog
pocketai-cdn/video-catalog.json	CDN - video generation model catalog
pocketai-cdn/voice-catalog.json	CDN - voice model catalog (STT, TTS, VAD)
pocketai-cdn/audio-catalog.json	CDN - upstream-linked music and sound generation catalog
pocketai-cdn/gated-catalog.json	CDN - gated/uncensored models (non-app-store builds only)
pocketai-cdn/catalog-versions.json	CDN - version numbers for update checks
pocketai-mono/packages/core/src/catalog-manifest.json	Local bundled copy - embedded in the app's core package. Structure: { "version": 1, "models": [...] } wrapping the same data as catalog.json

When adding or updating models, edit the appropriate CDN catalog, bump its version, then run npm run sync:catalogs from pocketai-mono.

App Store vs Website Builds

Gated models (gated-catalog.json) are completely stripped from app store binaries at compile time. The defaults are all "website" (gated content visible) — flip the flag only when building for app stores.

Desktop (Next.js / Tauri)

# Website build (default) — gated content included
npm run build

# App store build — gated content stripped from JS bundle
NEXT_PUBLIC_FOR_APP_STORES=true npm run build

Config: apps/app/app/app-config.ts reads process.env.NEXT_PUBLIC_FOR_APP_STORES. Next.js inlines this at build time, so the minifier removes the dead branch and the gated-catalog.json URL never appears in the bundle.

Android

# Website build (default)
./gradlew assembleRelease

# App store build — R8 strips dead branch in release builds
./gradlew assembleRelease -PFOR_APP_STORES=true

Or set in gradle.properties:

FOR_APP_STORES=true

Config: app/build.gradle.kts sets BuildConfig.FOR_APP_STORES. R8 (isMinifyEnabled = true) eliminates dead code in release builds.

iOS

Website build (default):  No extra flags needed.
App store build:          Xcode → Target → Build Settings →
                          "Active Compilation Conditions" → add FOR_APP_STORES

Config: PocketAI/Services/AppConfig.swift uses #if FOR_APP_STORES compiler directives. Code inside #if !FOR_APP_STORES ... #endif is not compiled into the binary at all.

Prebuilt mesh wheels

The wheels/ tree hosts prebuilt CUDA C++ extensions used by the 3D mesh generation backends (TRELLIS, Hunyuan3D). End users on a fresh Windows 11 machine without Visual Studio / CUDA toolkit cannot compile these from source, so per-backend venvs install them from the CDN via --find-links.

Path layout

cdn.pocketaihub.com/wheels/<variant>/<package>/<file>.whl
cdn.pocketaihub.com/wheels/<variant>/SHA256SUMS

The <variant> tag encodes torch + CUDA: pt<torch_major><minor>cu<cuda>. Current variant: pt24cu121 = torch 2.4.1 + CUDA 12.1.

Hosted packages

Package	Upstream	Used by
nvdiffrast	NVlabs/nvdiffrast	TRELLIS, Hunyuan3D
custom-rasterizer	Tencent-Hunyuan/Hunyuan3D-2.1 (hy3dpaint/custom_rasterizer)	Hunyuan3D
differentiable-renderer	Tencent-Hunyuan/Hunyuan3D-2.1 (hy3dpaint/DifferentiableRenderer)	Hunyuan3D

NOT hosted: diff-gaussian-rasterization. The Inria rasterizer (graphdeco-inria/diff-gaussian-rasterization) is non-commercial license and we will not redistribute binaries. TRELLIS-mesh-only does not require it; only the splat/NeRF path (intentionally excluded from PocketAI) does.

Install usage

pip install --find-links https://cdn.pocketaihub.com/wheels/pt24cu121/ nvdiffrast

Or pin exact URLs in a per-backend requirements.txt (Phase 4 subagents will wire this up when they author each backend):

# For packages with CUDA build deps baked in:
nvdiffrast @ https://cdn.pocketaihub.com/wheels/pt24cu121/nvdiffrast/nvdiffrast-0.3.3+pt24cu121-cp311-cp311-linux_x86_64.whl ; sys_platform == 'linux'
nvdiffrast @ https://cdn.pocketaihub.com/wheels/pt24cu121/nvdiffrast/nvdiffrast-0.3.3+pt24cu121-cp311-cp311-win_amd64.whl ; sys_platform == 'win32'

Filenames follow PEP 427 with the +<variant> local-version segment: {package}-{version}+pt24cu121-cp311-cp311-{linux_x86_64|win_amd64}.whl.

How to add a new wheel

Wheels are produced by the build-mesh-wheels.yml workflow in the pocketai-mono repo and committed here automatically.

• Manual run: Actions → "Build mesh wheels" → Run workflow.
• Tag-triggered: push a tag matching mesh-wheels-v*.

The workflow's publish job clones this repo over SSH using the POCKETAI_CDN_DEPLOY_KEY secret (configured on the pocketai-mono repo), copies wheels into wheels/<variant>/<package>/, regenerates the PEP 503 simple-index index.html pages, refreshes SHA256SUMS, and pushes one commit per new wheel file plus a final index-refresh commit.

To add a new variant (e.g. pt24cu124): update VARIANT / CUDA_VERSION / TORCH_VERSION env in the workflow and add a matching directory here.

Usage in Apps

The desktop/web app references the CDN base URL in apps/app/app/api-client.ts:

const CDN_URL = process.env.NEXT_PUBLIC_CDN_URL || "https://cdn.pocketaihub.com";