News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
MiniCPM

MiniCPM

Efficient small language models for end-side devices

MiniCPM is a family of small language models built by OpenBMB to run efficiently on local devices. Released from 2024, it uses scalable training strategies to match larger models while keeping resource use low.

Share on WhatsApp

  • MiniCPM4-8B trained on 8 trillion tokens
  • Supports 128K context window via YaRN
  • 7x faster decoding on Jetson AGX Orin
  • Released under Apache-2.0 licence
  • Designed for efficient end-side deployment

What it is

MiniCPM is a series of small language models designed for deployment on end-side devices. Developed by OpenBMB, the project aims to provide high performance with low computational cost. The latest version, MiniCPM4, includes an 8-billion parameter model trained on eight trillion tokens. It focuses on efficiency in architecture, data, algorithms and inference systems. The models are intended for users who need capable AI without relying on large cloud servers.

How it came about

The project was introduced in April 2024 with a paper detailing its scalable training strategies. The initial release included 1.2-billion and 2.4-billion parameter variants. These early models showed performance comparable to larger seven to thirteen billion parameter systems. The team used extensive experiments to find stable scaling methods. In June 2025, the MiniCPM4 series was released, marking a significant update focused on ultra-efficiency for local hardware.

Inside the model

The models use a decoder-only transformer architecture. Training employs a Warmup-Stable-Decay learning rate scheduler, which allows for continuous training and domain adaptation. This approach helps study data-model scaling laws without extensive retraining. MiniCPM4 introduces systematic innovations across four areas: model structure, training data, algorithms and inference systems. It supports long context windows up to 128K tokens using YaRN technology for length extension.

What it can do

MiniCPM models perform well in their size category, matching capabilities of much larger models. The 8-billion parameter version shows strong results in long text processing. On Jetson AGX Orin chips, it achieves roughly seven times faster decoding speed than similar models. It handles needle-in-a-haystack tasks effectively within its 128K context window. The smaller 0.5-billion variant also claims best-in-class status for its scale.

How to use it

The models are released under the Apache-2.0 licence, allowing open use and modification. Weights are available publicly through platforms like Hugging Face. Users can run the models locally using tools such as vLLM. The developers note that special tokens must be handled explicitly in some API configurations to ensure correct formatting. Technical reports and code are provided to support implementation.

Safety and testing

The developers state that the model generates content based on learned text patterns. It does not have personal opinions or value judgments. Generated output does not reflect the views of the creators. Users are advised to take full responsibility for evaluating and verifying any content produced by the system. No specific red-teaming or safety evaluation details are provided in the available materials.

Why it matters

MiniCPM highlights the potential of small language models as efficient alternatives to massive systems. Its training strategies offer insights into compute-optimal data-model ratios. By demonstrating high performance on end-side chips, it supports the trend towards local AI deployment. The open release encourages further research and application in resource-constrained environments.

Questions people ask about MiniCPM

What is MiniCPM used for?

It is designed for running on local devices like laptops or edge hardware. It provides general language capabilities such as text generation and analysis without needing cloud connectivity.

How large is the MiniCPM4 model?

The flagship MiniCPM4-8B model has eight billion parameters. There is also a smaller 0.5-billion parameter version available for even more constrained devices.

Is MiniCPM free to use?

Yes, the models are released under the Apache-2.0 licence. This permits open use, modification and distribution, making them accessible for both research and commercial applications.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on MiniCPM yet, so the facts on this page come from the sources listed above: huggingface.co/openbmb/MiniCPM4-8B, arxiv.org/abs/2404.06395. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.