CubicLM

A cross-platform AI chat application with local on-device inference and multi-provider cloud AI support.

Runs LLMs directly on your Android device via GPU-accelerated llama.cpp and Google's LiteRT-LM runtime, with an optional built-in OpenAI-compatible API server.

Features

Local AI inference

- LLM inference via llama.cpp (GGUF models) with GPU acceleration (Vulkan / OpenCL) - LiteRT-LM inference via Google's LiteRT-LM runtime (.litertlm models) - Stable Diffusion 1.5 on-device image generation (safetensors) - Vision models — Qwen2-VL-2B, Gemma 4 E2B/E4B for image understanding - Streaming token generation with real-time tokens-per-second display - GPU crash recovery with automatic CPU fallback - Device-tier auto-configuration based on detected RAM - Real hardware identification (SoC and GPU renderer)

Inference parameters

- Auto Tune mode with optimal context and output limits per device - Manual mode with context window up to 1M tokens and output up to 128K tokens - Inference temperature control - Local sampling controls (Top-P, Top-K, repeat penalty) - RAM guard with safety clamping - Compute backend toggle (CPU / Vulkan / OpenCL)

Cloud AI providers

- OpenRouter, Hugging Face, xKiro, TokenRouter, AgentRouter, OrcaRouter, APInex - OpenAI, Anthropic, Google Gemini, DeepSeek, Z.AI, Groq, Mistral AI - Together AI, xAI Grok, Perplexity, Cerebras, Fireworks AI, Cohere, NVIDIA NIM - Stability AI for cloud image generation - Custom OpenAI-compatible endpoints with multiple profiles

Chat features

- Multi-session chat with history - Full-text searchable sidebar - Message actions — copy, share, regenerate, branch, edit with revision history - Prompt templates — 6 built-in plus custom - Multi-select bulk operations - Per-chat model pin - Side-by-side model comparison - Battle Arena — up to 4 cloud models race one prompt - Chat labels and hidden chats - Whole-chat PDF export

Web access

- Fetch live web content from links in messages - Clean text extraction for model context - Visible source chips with favicon and title

Skills

- Offline instruction extensions (markdown) - Intelligent per-prompt activation - 5 bundled starter skills - Import from file, URL, or Anthropic skills browser

Custom MCP Server

- Single remote connection (Streamable HTTP / SSE) - Optional bearer token authentication - Tool preview and ask-before-run dialog

Built-in OpenAI-compatible API server

- Expose local models on port 8080 - Optional API key authentication - Rate limiting

Disclaimer

CubicLM is an independent, open-source project. Not affiliated with, endorsed by, or connected to any AI model provider or third-party services. All trademarks belong to their respective owners.

License

MIT