面壁开源 MiniCPM 端侧模型

2024 年 2 月 1 日，面壁智能与 OpenBMB 发布 MiniCPM-2B。模型卡给出的主体语言模型规模约为 24 亿非词嵌入参数。这个数字放在当时的开放模型列表里并不显眼：市场习惯用 7B、13B、70B 判断能力档位。MiniCPM 选择的展示却不是再画一条更长的参数柱，而是把 Int4 量化版本装进手机，让文本在设备上直接生成。

手机改变了问题的顺序。云端服务先问吞吐和集群成本，端侧模型先面对内存、电量、温度与首 token 延迟。权重必须被压缩，运行时必须适配移动芯片，持续生成又不能让设备迅速过热。官方材料称，量化后的 MiniCPM 可以在手机上推理，流式输出速度略高于人类说话速度；这是一项发布方演示与测量，不代表所有手机和所有上下文都能达到同样表现，却足以证明推理不必总要往返数据中心。

团队还把小规模与可训练性绑在一起。模型卡给出的路径包括 SFT、DPO 和 Int4 版本，并称一张 1080 或 2080 可进行参数高效微调，3090 或 4090 可尝试全参数微调。权重、训练检查点和大部分非专有数据按研究及有限商用条件发布。对个人开发者与小团队，这意味着实验的起点不再是一份 API 账单：可以下载模型、改动它，并在自己的硬件边界里看见失败。

官方基准称 SFT 版本在综合评测上接近 Mistral-7B，并在中文、数学和代码任务上表现突出。这些数字来自项目方，不能被写成独立结论。模型卡自己也列出限制：小容量会带来幻觉，结果更受提示词影响，知识记忆不够准确。小模型的价值不在于神奇地抹平参数差距，而在于用能力损失交换可控延迟、本地数据和更低的部署门槛。

MiniCPM-V 随后把图像理解加入同一端侧路线，并展示多模态模型在手机上运行。它是系列延伸，不应被倒写成 2 月 1 日首版语言模型已经具备的全部能力。更重要的连续性在于，后续版本仍围绕量化、长上下文、视觉和边缘芯片调整：端侧不是一次营销演示，而成为模型家族必须持续回答的工程约束。

MiniCPM 没有把手机变成数据中心，也没有让云模型失去意义。它做的是把“在哪里计算”变成能力定义的一部分。同一个回答，如果必须上传私人数据、等待网络并支付远程推理费用，与它在本机离线生成，并不是同一种产品。24 亿参数的意义不在参数榜上越级夺冠，而在于它让模型真正住进了要被使用的设备。

ModelBest and OpenBMB released MiniCPM-2B on 1 February 2024. Its model card described about 2.4 billion non-embedding parameters in the main language model. That number did not dominate an open-model list accustomed to 7B, 13B, and 70B tiers. MiniCPM’s more consequential demonstration was elsewhere: an Int4-quantized version running on a phone, generating text on the device itself.

A phone changes the order of the questions. Cloud serving begins with cluster throughput and fleet cost. On-device inference meets memory, battery, temperature, and time to first token before anything else. Weights must be compressed, runtimes adapted to mobile chips, and sustained generation prevented from exhausting the thermal budget. The project reported that quantized MiniCPM could infer on smartphones at a streaming speed slightly faster than ordinary human speech. That was a publisher demonstration, not a guarantee for every phone and context, but it established that an answer did not always need a round trip to a data center.

The team connected small scale with trainability as well. The model card offered SFT, DPO, and Int4 variants, and said parameter-efficient fine-tuning could be performed on a 1080 or 2080 GPU, while a 3090 or 4090 could support full-parameter work. Weights, training checkpoints, and much of the non-proprietary data were released under terms for research and limited commercial use. For individuals and small teams, experimentation no longer had to begin with an API bill. They could download the model, change it, and observe failure inside hardware they controlled.

Official evaluations said the SFT model approached Mistral-7B on broad benchmarks and was stronger in Chinese, mathematics, and code. Those were project results, not an independent verdict. The model card also documented its own limits: small capacity increased hallucination, made outputs more sensitive to prompts, and weakened accurate knowledge storage. The value of a small model was not that parameter differences had magically disappeared. It was the ability to trade some capability for predictable latency, local data, and a lower deployment threshold.

MiniCPM-V soon extended the same on-device direction to image understanding and demonstrated multimodal inference on phones. That was a continuation of the family, not a capability that should be inserted retroactively into the first language-model release. The more important continuity was the engineering constraint itself. Later versions kept returning to quantization, long context, vision, and edge chips. On-device operation became a requirement the family had to answer repeatedly rather than a single marketing demo.

MiniCPM did not turn a phone into a data center, and it did not make cloud models irrelevant. It made where the computation happens part of the capability definition. An answer that requires uploading private data, waiting for a network, and paying a remote inference bill is a different product from one generated offline on the device. The significance of 2.4 billion parameters was not an upset victory on a size leaderboard. It was that the model could live inside the machine where it would actually be used.