MiniCPM5-1B is a dense 1B open model built for on-device and local deployment. It supports 131K context, Think / No Think modes, tool calling, GGUF and MLX formats, major inference backends, and even powers an offline desktop pet.

MiniCPM-V 4.6 is an open MLLM for image and video understanding on phones and consumer hardware, with mixed 4x/16x visual token compression, iOS/Android/HarmonyOS demos, and support for vLLM, SGLang, llama.cpp, and Ollama.

VoxCPM2 is a 2B open-source TTS model with 30-language support, 48kHz output, voice design from text alone, controllable voice cloning, and real-time streaming fast enough for production voice workflows.

A 9B omni-modal model that sees, listens, and speaks simultaneously. Features full-duplex streaming (no turn-taking lag) and proactive interaction. Outperforms GPT-4o on vision benchmarks. Runs locally via llama.cpp & Ollama.

MiniCPM 4.1 is a new 8B open-source model designed for the edge. Featuring a novel trainable sparse attention architecture, it brings efficient "deep thinking" and long-context capabilities to on-device AI, achieving state-of-the-art performance for its size.

MiniCPM-V 4.5 is a new 8B open-source MLLM that delivers GPT-4o level performance on your phone. It excels at image, video, and document understanding, beating top proprietary models on key benchmarks like OCRBench.

VoxCPM is a new open-source, tokenizer-free TTS model. By modeling speech in a continuous space, it overcomes the limitations of discrete tokens to deliver highly expressive, context-aware speech generation and incredibly realistic zero-shot voice cloning.

MiniCPM 4.0 is a family of ultra-efficient, open-source models for on-device AI. Offers significant speed-ups on edge chips, strong performance, and includes highly quantized BitCPM versions.