MiniCPM (OpenBMB / ModelBest 面壁智能)
Small on-device LLMs and multimodal models that outrun bigger rivals
Overview
MiniCPM is OpenBMB/ModelBest's (面壁智能) line of small open-weight language and multimodal models — MiniCPM-V for vision, MiniCPM-o for real-time omnimodal streaming — tuned to run on phones and edge hardware, developed out of Tsinghua's NLP lab.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 3 Sep 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
What you can produce with MiniCPM (OpenBMB / ModelBest 面壁智能)
- MiniCPM-V vision-language models for on-device image/video understanding
- MiniCPM-o omnimodal models with real-time streaming audio/video
- MiniCPM5-1B dense on-device LLM
- 96x video-token compression in MiniCPM-V 4.5
- Apache 2.0 code license (model weights under OpenBMB community license)
- Free commercial use under registration (up to 5,000 edge units / 1M DAU)
MiniCPM's whole argument is punching above its weight class: MiniCPM-o 4.5 (9B parameters) scores 78.2 on OpenCompass's eight-benchmark suite, ahead of GPT-4o and Gemini 2.0 Pro, and MiniCPM-V 4.6 beats Qwen3.5-0.8B's intelligence score while using 19x fewer tokens — genuinely useful if you're deploying vision-language models on hardware that can't run a 70B model. The February 2026 China Telecom-led financing (multi-hundred-million yuan) and the Ruixinwei deal to build MiniCPM-o4.5-powered AI Box units for car cockpits show it's moving from research demo to embedded product.
It's not a hosted SaaS: there's no first-party chat UI or managed API, so you're self-hosting or going through a third-party inference provider. Code is Apache 2.0, but model weights carry OpenBMB's own community license — free for research and most commercial use, but usage caps (5,000 edge units or 1M DAU) require registering, and above that, direct authorization.
What people say
MiniCPM comes out of THUNLP — Tsinghua University's NLP lab — with ModelBest (面壁智能) as the commercial arm, founded in August 2022 by Liu Zhiyuan and other lab affiliates. The pitch has stayed consistent across versions: match or beat much larger models on a fraction of the parameters, small enough to run on a phone.
The benchmark numbers back that up more than most 'small model' claims do. MiniCPM-o 4.5, a 9B end-to-end omnimodal model built on SigLip2, Whisper-medium, CosyVoice2 and Qwen3-8B, averages 78.2 across OpenCompass's eight-benchmark suite — ahead of GPT-4o and Gemini 2.0 Pro on that measure, and close to Gemini 2.5 Flash. MiniCPM-V 4.5 hits 77.0 on the same suite at 8B parameters and can compress six 448x448 video frames into 64 tokens, a 96x reduction versus the roughly 1,536 tokens most MLLMs need for the same input — the reason it keeps showing up in edge-deployment shortlists. MiniCPM-V 4.6, aimed even further down the hardware ladder (SigLip2-400M plus a 0.8B LLM), scores 13 on the Artificial Analysis Intelligence Index against Qwen3.5-0.8B's 10, using 19x fewer tokens to get there.
Money has followed the technical story. ModelBest closed a multi-hundred-million-yuan round in February 2026 led by China Telecom, with Zhongxin Jingshi and Zhongxin Private Equity also participating — China Telecom's stated plan is to pair its cloud/network/edge infrastructure with ModelBest's vertical experience in judicial, automotive and education deployments. In late April 2026, ModelBest and Ruixinwei Electronics signed a deal to build AI Box products for car cockpits on top of MiniCPM-o4.5. Earlier backers include Hillhouse Capital and CITIC Securities.
The catch for anyone evaluating this as a product rather than a research release: there isn't one, in the SaaS sense. No hosted chat app, no managed API with a dashboard and a credit card — you pull weights from Hugging Face or GitHub and either self-host or route through a third-party inference host (SiliconFlow among them, as it happens). Licensing is split: code under Apache 2.0, but the model weights carry OpenBMB's own community license, free for research and for commercial use under registration up to 5,000 edge-device units or 1 million daily active users; past that, you need to email OpenBMB directly for authorization.
Summary of public user & expert reviews, compiled by RECATOOLS.
About this listing
This entry was compiled from publicly available data including MiniCPM (OpenBMB / ModelBest 面壁智能)'s official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with MiniCPM (OpenBMB / ModelBest 面壁智能) unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to MiniCPM (OpenBMB / ModelBest 面壁智能) directly →
Spotted something out of date? Suggest an update →
Alternatives to MiniCPM (OpenBMB / ModelBest 面壁智能)
More in LLMs & Chat
