™️
Small but powerful, hybrid thinking with 512K context
Foundation Model MiniCPM
- 2B
- 1B
- 8B
- 4B
- 0.5B
- 2.4B
- 1.2B
HuggingFace 
SMALL SIZE, SUPER POWER
Edge Model for Everyone, Everyday, Everywhere
MiniCPM
InsidePhones
MiniCPM
InsideAIPC
MiniCPM
InsideIntelligent
Cabins
MiniCPM
InsideEmbodied
Robots
MiniCPM
InsideWearable
Devices
Flagship On-device Model Series
The MiniCPM® edge model series is a world-leading, lightweight, and high-performance LLM. Since its release in February 2024, it has been widely tested and acclaimed by the global open-source community for its "achieving more with less" efficiency and outstanding on-device performance. It has repeatedly topped GitHub and Hugging Face trending charts, becoming one of the most popular LLMs on Hugging Face in 2024. The MiniCPM® has partnered with industry benchmark leaders, emerging as an indispensable player in driving innovation across sectors such as AIPC, AI phones, intelligent cabins, and embodied robots.
A flagship on-device language model with 8B sparse and 0.5B lightweight variants, up to 220× faster inference, 128K long context, benchmark-level quality with only 22% training cost, and a 2GB quantized footprint.
ModelBest’s flagship multimodal series. MiniCPM-O is a full-duplex omni model supporting streaming video, audio, text input, and speech output; MiniCPM-V is an efficient on-device vision model that runs on phones with 6GB memory across iOS, Android, and HarmonyOS.
ModelBest’s first embodied intelligence series. Its 1.5B VLA model delivers performance comparable to 3–4B models with one-minute native memory and streaming inference, while its tracking model supports fully offline visual tracking.
An end-to-end tokenizer-free speech synthesis model covering 30 languages and 9 Chinese dialects. It supports voice design from text descriptions, zero-shot voice cloning, and 48kHz CD-quality output under the Apache 2.0 license.
Flagship On-device Model Series
The MiniCPM® edge model series is a world-leading, lightweight, and high-performance LLM. Since its release in February 2024, it has been widely tested and acclaimed by the global open-source community for its "achieving more with less" efficiency and outstanding on-device performance. It has repeatedly topped GitHub and Hugging Face trending charts, becoming one of the most popular LLMs on Hugging Face in 2024. The MiniCPM® has partnered with industry benchmark leaders, emerging as an indispensable player in driving innovation across sectors such as AIPC, AI phones, intelligent cabins, and embodied robots.
A production-grade LLM training framework written autonomously by AI. It matched NVIDIA Megatron-LM in eight hours, surpassed it within 1.5–2 days, and improved measured H100 MFU by about 8%–10%.
ForgeStencil
GitHub A closed-loop AI system for automated Stencil research and deployment. Its dual-agent architecture refactored over 100 real industrial and scientific applications in one week with zero human intervention.
Global Partner
Purely On-Device! Superior Performance! Full-Scope Scenarios!
The First Purely On-Device Intelligent Assistant
On-Device Native
Personalized for All Scenarios
Chip-Level Fit
AI Native OA
PLAY
Large Model
Agent
Infra
Ultra Alignment
Others
Efficiency FirstWe believe the best model is the one with superior power, faster speed and lower costEfficiency comes from mastering the science of large language models (LLMs), with knowledge density as the key principle.
As knowledge density grows, it becomes a core competitive advantage, unlocking vast potential for edge intelligence and applications.
Modelbest LawMoore’s Law
Model capability density increases exponentially over time, with the number of parameters required to reach a certain intelligence level halving every 3.3 months.Capability density: The ratio of effective parameter size to actual parameter size. Effective parameter size refers to the minimum number of parameters required for the reference model (e.g., MiniCPM) to achieve performance equivalent to the given target model.
A G I F O R L I V E S