OpenBMB发布MiniCPM5-2B：端侧小模型实现复杂Agent推理与工
Insane progress for small language models!
2B参数模型在端侧实现复杂多步Agent推理是重大突破，适合关注边缘计算与轻量化AI的工程团队参考其训练范式与部署方案。
Insane progress for small language models!
小型语言模型取得了惊人的进展！
MiniCPM5-2B is a dense 2B-parameter model by OpenBMB from China that's built for reasoning, coding, and tool use on resource-constrained hardware.
MiniCPM5-2B 是中国 OpenBMB 推出的一款拥有 20 亿参数的密集模型，专为在资源受限的硬件上进行推理、编码和工具使用而构建。
The model specifically excels at coding and tool calling, two capabilities central to the shift from on-device LLMs to on-device agents.
该模型在编码和工具调用方面表现尤为出色，这两项能力是从端侧大语言模型向端侧智能体转变的核心。
Instead of only answering prompts, it can use tools, generate code, carry information between steps, and complete multi-step tasks.
它不仅能回答提示词，还能使用工具、生成代码、在步骤间传递信息，并完成多步任务。
I ran it 100% locally and connected it to a small investigation agent with one request:
我在本地完全运行了该模型，并将其连接到一个小型调查智能体，发出如下请求：
> Revenue dropped last week. Investigate what happened, quantify the impact, identify the likely cause, and produce an incident report with supporting evidence.
> 上周收入下降。请调查发生了什么，量化影响，确定可能的原因，并生成一份附带支持证据的事故报告。
The evidence was spread across orders, traffic, payments, refunds, and deployment logs.
证据分散在订单、流量、支付、退款和部署日志中。
The model inspected the files, wrote its own queries, analyzed the intermediate results, and decided what to investigate next.
模型检查了这些文件，编写了自己的查询语句，分析了中间结果，并决定下一步调查什么。
Each tool result informed the next action, so the final report depended on the model maintaining a coherent investigation across the complete trajectory.
每个工具的结果都影响了后续行动，因此最终报告取决于模型在整个过程中保持连贯的调查逻辑。
The recording shows the actual task from beginning to end. It starts with the revenue question, follows the tool calls and supporting evidence, and ends with a quantified diagnosis and incident report.
视频记录了从开始到结束的实际任务过程。它以收入问题为起点，跟随工具调用和支持证据，最终以量化的诊断结论和事故报告结束。
The data, tool execution, and model inference all remained on my machine.
数据、工具执行和模型推理均在我的机器上完成。
These capabilities were optimized through Agentic Pre-training, SFT, and large-scale RL. They do not come entirely from an application-level agent framework.
这些能力是通过代理式预训练（Agentic Pre-training）、监督微调（SFT）和大规模强化学习（RL）优化而来的。它们并非完全来自应用层的智能体框架。
MiniCPM5-2B supports SGLang, vLLM, llama(.)cpp, Ollama, iOS, Android, and HarmonyOS.
MiniCPM5-2B 支持 SGLang、vLLM、llama.cpp、Ollama、iOS、Android 和 HarmonyOS。
OpenBMB has also released the model weights and parts of the training recipes and data resources behind it.
OpenBMB 还发布了该模型的权重，以及部分背后的训练配方和数据资源。
Download MiniCPM5-2B: https://huggingface.openbmb.cn/model/openbmb/MiniCPM5-2B?from=_avichawla
下载 MiniCPM5-2B：https://huggingface.openbmb.cn/model/openbmb/MiniCPM5-2B?from=_avichawla
A 2B model can now maintain enough state to coordinate tools and complete a useful investigation on local hardware.
如今，一个 20 亿参数的模型已能在本地硬件上维持足够的状态以协调工具，并完成有用的调查任务。
更进一步：量化金融体系
看懂新闻只是起点——沿量化金融路径，把它变成能交付的工程能力