Open-source 48kHz TTS with voice design and cloning Open Source • Artificial Intelligence • Audio 4 114 Lightfield AI-native CRM that builds itself and does work for you Sales • Artificial Intelligence • CRM
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.
MiniCPM-SALA variant supports million-token context with sparse-and-linear hybrid attention architecture. MiniCPM5-1B includes native long-context support, enabling efficient processing of extended documents on limited hardware.
MiniCPM provides single-page cookbooks for major fine-tuning frameworks paired with Agent Skills. The repository includes training recipes, quantization options (GPTQ, AWQ, GGUF), and access to training datasets (UltraData-SFT, Ultra-FineWeb) for reproducible customization.
MiniCPM is released on HuggingFace and ModelScope with full technical reports, training recipes, and implementation details. Multiple quantized variants (BF16, GGUF, MLX) and deployment guides enable rapid integration into research or production pipelines.
MiniCPM5-1B is a 1B-parameter dense transformer optimized for on-device deployment with minimal footprint. It reaches 1B-class open-source SOTA performance while maintaining compatibility with edge chips, enabling local inference without cloud dependency.
MiniCPM5-1B includes built-in hybrid reasoning via <think> chat template and demonstrates strong performance in tool use, code generation, and agentic benchmarks. The model supports both fast assistant and deliberate reasoner modes from a single checkpoint.