Changelog
Text Models (MiniCPM Text Series)
MiniCPM5-2B
MiniCPM5-2B released: a 2B dense model for on-device and resource-constrained scenarios, with state-of-the-art performance among models under 4B parameters.
MiniCPM5-1B
MiniCPM5-1B released! A 1B-parameter on-device SOTA model with hybrid reasoning (Think / No Think), including Agent Skills for deployment and fine-tuning.
MiniCPM-SALA
MiniCPM-SALA released with hybrid sparse–linear attention, million-token context support, and efficient inference.
InfLLM-V2 Paper
The InfLLM-V2 paper is out! Train sparse-attention models with only 5B long-context tokens.
MiniCPM4.1-8B
MiniCPM4.1-8B released: a trainable sparse-attention model supporting hybrid reasoning.
MiniCPM4
The MiniCPM4 series is out! It retains best-in-class performance at its size while improving efficiency, delivering over 5× faster generation on typical on-device chips.
Multimodal Models (MiniCPM-V Series)
MiniCPM-OCR
MiniCPM-OCR released with MiniCPM-OCR 1.0, a 1.3B parsing model, and MiniCPM-OCR Jev, a 2B System-1 decision model. MiniCPM-OCR 1.0 supports efficient document parsing and text recognition across scans, photographs, curved, skewed and recaptured documents. Jev assesses text, formula and table parsing results and returns calibrated probabilities to help evaluate quality and prioritize review.
MiniCPM-V 4.7
MiniCPM-V 4.7 released in 35B-A3B and 1B variants, with further improvements in multimodal understanding and inference efficiency. The 35B-A3B variant is the series' first MoE model, supporting smooth inference on phones and other local devices while preserving model capabilities.
MiniCPM-V 4.6 API Service
API services for MiniCPM-V 4.6 and MiniCPM-o 4.5 are available. MiniCPM-V 4.6 offers a public, free API key!
MiniCPM-V 4.6
MiniCPM-V 4.6 is open source: an efficient 1.3B-parameter multimodal model with hybrid 4×/16× visual token compression, approximately 1.5× the throughput of Qwen3.5-0.8B, and deployment on iOS, Android, and HarmonyOS.
MiniCPM-V 4.5
MiniCPM-V 4.5 is open source, surpassing GPT-4o-latest, Gemini-2.0 Pro, and Qwen2.5-VL 72B.
MiniCPM-V 4.0
MiniCPM-V 4.0 is open source, surpassing GPT-4.1-mini in image understanding with improved efficiency. The iOS app is also open source.
Omnimodal Models (MiniCPM-o Series)
MiniCPM-o 4.5 Technical Report
The MiniCPM-o 4.5 technical report is out, covering the core technologies behind full-duplex, real-time omnimodal interaction.
MiniCPM-o 4.5 Real-time Web Demo
The MiniCPM-o 4.5 real-time Web Demo is available, supporting local deployment on Mac and GPU.
MiniCPM-o 4.5
MiniCPM-o 4.5 is open source! This 9B-parameter full-duplex multimodal model offers vision and speech capabilities close to Gemini 2.5 Flash, enabling real-time interaction while seeing, listening, and speaking simultaneously.
MiniCPM-o 2.6 Technical Report
The MiniCPM-o 2.6 technical report is out.
MiniCPM-o 2.6
MiniCPM-o 2.6 is open source, with vision, speech, and multimodal streaming capabilities at the GPT-4o-202405 level, plus new features.
Speech Models (VoxCPM Series)
VoxCPM2
VoxCPM2 released! A 2B-parameter speech model supporting 30 languages, voice design, controllable voice cloning, and 48 kHz audio output.
VoxCPM1.5
VoxCPM1.5 is open source, supporting SFT and LoRA fine-tuning and reaching the top of GitHub Trending!
VoxCPM-0.5B
VoxCPM-0.5B is open source: the first on-device speech model, reaching the top of Hugging Face Trending!