Most people still assume that "real" artificial intelligence lives in a data center — racks of GPUs humming behind a cloud API. A small Beijing lab is quietly arguing the opposite. Its models are small enough to run on the phone already in your pocket, yet they read documents, describe scenes, and hold a conversation without ever touching a server.
The lab is ModelBest (面壁智能), working with Tsinghua University's NLP group under the open-source banner OpenBMB (面壁智能与清华大学自然语言处理实验室). Their MiniCPM family is the clearest example in China of a deliberate bet: make the model small, make it efficient, and put it where the user actually is.
What "end-side" really means
In China's AI scene the term is 端侧大模型 (end-side large model) — a model that runs on the device itself rather than calling a remote server. The implications are not subtle:
- No cloud bill. Once the weights are on the device, each inference costs almost nothing.
- No round trip. Answers appear as fast as the device can compute, often faster than a person speaks.
- Privacy by default. Your photos, messages, and documents never leave the hardware.
- Offline works. A dead connection does not silence the assistant.
This is not a toy category. Shipments of AI phones and AI PCs in China have pushed vendors to demand models that fit in a few gigabytes of memory.
The numbers behind the small model
The original MiniCPM, released in 2024, carried just 2.4 billion parameters (excluding embeddings) yet scored close to Mistral-7B on standard benchmarks and beat much larger models on Chinese, math, and coding tasks after fine-tuning. The team reported that its streaming output on a phone ran faster than human speech — a claim you can actually test by loading it locally.
The multimodal branch, MiniCPM-V, is where the story gets interesting:
- MiniCPM-V 2.0 (2.4B) reached scene-text OCR performance comparable to Google's Gemini Pro on several benchmarks, despite being a fraction of the size.
- The latest MiniCPM-V line runs a 1.3B vision-language model on phones with as little as 6GB of RAM, covering iOS, Android, and HarmonyOS (鸿蒙).
- Quantized versions need only about 2GB of memory, and the family advertises up to 220× inference speedups versus naive deployment.
- Context length reaches 128K tokens, enough to hold a short book in memory.
From vision to "everything at once"
ModelBest did not stop at images. The MiniCPM-O branch is an end-to-end, full-duplex, full-modal model around 9B parameters that takes video, audio, and text as input and streams speech back out — the kind of always-listening, always-seeing interaction that usually requires a server farm. The company positions it as a mobile counterpart to Google's Gemini 2.5 Flash, a comparison that is marketing, not an independent benchmark, but the architectural direction is real.
There is also MiniCPM-Robot (1.5B), a vision-language-action model for embodied (具身) robotics, and VoxCPM, a speech-synthesis model. The throughline is the same: shrink the model until it fits the edge, then make the edge do more.
Why smaller can win
The instinct that "bigger is better" hides a cost curve. A 70B or 600B model needs expensive hardware, constant power, and a network connection. A 1–9B model needs a phone. For the billions of devices already shipped, that difference is the difference between a feature and a fantasy.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the end-side race is less about bragging rights on leaderboards and more about who can make AI disappear into everyday hardware — "the model that wins the pocket is the model that wins the market."
The honest trade-offs
Small models hallucinate more. The MiniCPM team itself notes that, because of limited capacity, outputs are sensitive to prompts and knowledge recall is not always accurate. They are not replacing frontier models for deep reasoning or open-ended research. They are replacing the cloud for the boring, private, always-on tasks: reading a receipt, captioning a photo, translating a menu, controlling a device by voice.
That is a large market, and it is one where China's dense hardware ecosystem — phones, cars, wearables, home appliances — gives local labs a natural home field.
Honest limitations
Figures here come from ModelBest's official site and the project's GitHub README; several performance claims (220× speedup, Gemini 2.5 Flash comparison) are company-disclosed and have not been independently audited. Parameter counts and the fact of on-device deployment are well documented. Benchmark numbers from earlier MiniCPM versions (2.0, original 2.4B) are reproducible from released weights. This article does not cover pricing, enterprise deals, or the separate question of how these models perform against the very latest 2026 flagship models, which would require fresh testing.
What readers can do now
- Try it yourself. Pull a MiniCPM-V or the 1.3B vision model from Hugging Face or ModelScope and run it through Ollama or llama.cpp on a laptop — no API key, no bill.
- Audit the claim. If a vendor says "runs on-device," ask for the quantized memory footprint and a live demo; a real end-side model should work with the phone in airplane mode.
- Match model to job. Use a 1–9B edge model for private, repetitive, latency-sensitive tasks, and reserve cloud frontier models for heavy reasoning — don't pay server costs for work a phone can do.