Summary
- The MiniCPM5-1B model achieves an average score of 42.57 on agentic and reasoning benchmarks, surpassing the closest competitor in the 1B-class category, which scored 35.61.
- This model is equipped with MCP and native tool calling capabilities, allowing it to operate local agent workflows on consumer devices without needing cloud access.
- Our evaluation revealed that while the model demonstrated strong conversational skills, it also generated a misleading chain-of-thought response and struggled with a basic logic question.
MiniCPM5-1B, developed by OpenBMB, is a one-billion-parameter model that is the newest addition to the MiniCPM series designed for on-device use. It supports native tool calling and the Model Context Protocol (MCP), fitting comfortably within a smartphone's memory, and outperforms other open-source models of similar size in benchmarks.
As the inaugural model in the MiniCPM5 series, it was specifically created for deployment on devices with limited resources. With its 1 billion parameters, it is considered small by today’s standards. A higher number of parameters typically indicates a more capable AI model.
In comparison, Google's Gemma 4 begins with 2 billion parameters and can expand up to 31 billion. Llama 4 Scout operates with 17 billion parameters. MiniCPM5-1B does not aim to compete with such models but instead focuses on maximizing performance with minimal resources.
Construction Details
The design of MiniCPM5-1B is based on MiniCPM4, as outlined in a technical report from the OpenBMB team at THUNLP, Tsinghua University, and ModelBest. A key innovation is the InfLLM v2 mechanism, which optimizes the processing of each token by analyzing fewer than 5% of surrounding tokens during long-context inference, significantly reducing computational demands without sacrificing accuracy. (In AI, a “token” refers to the smallest unit of information processed by the model.)
On the data front, the team employed UltraClean, a filtering method that enabled the model to achieve competitive performance using 8 trillion training tokens, in contrast to the 36 trillion utilized by Qwen 3. After training, reinforcement learning and efficient distillation techniques were applied, enhancing benchmark scores in math, coding, and instruction-following by 16 points while reducing excessive response lengths by 29 percentage points.
The model can manage a context window of 128K tokens, which translates to approximately 96,000 words of continuous text in one go. For a model with 1 billion parameters, this is a substantial capacity. It allows for persistent memory across lengthy roleplay sessions, comprehensive PDF summaries, or maintaining agent context without interruptions during tasks.
Why a Basic Agent Might Suffice
In our testing, MiniCPM5-1B was confirmed to support MCP and tool calls, placing it among a limited group of sub-2 billion-parameter models capable of executing genuine agentic workflows without relying on cloud services.
However, users will need to configure additional settings, which are detailed in the model's Github repository.
For practical applications, this means a local agent on a smartphone could check a calendar, search a local database, or access a web research MCP server—all while offline. As previously noted, running local AI is becoming more feasible than many expect, and the competition in this area is intensifying. Models specifically designed for mobile use are evolving into a legitimate product category rather than remaining solely a research curiosity.
There’s no need for OpenAI to check your calendar when a local agent can simply retrieve that information and inform you of your agenda for the day.
For lighter tasks and extended conversations, MiniCPM5-1B stands out in performance. Even though OpenBMB may not have considered it, the model's articulate style makes it suitable for local roleplay; with a context capacity of 128K, narratives can unfold through numerous exchanges without losing coherence.
It excels at managing small tasks such as reading notes, summarizing documents, and answering related questions, especially when used in conjunction with an MCP research server to fill knowledge gaps.
In the competitive landscape, it faces off against models like Alibaba's Qwen3-0.6B, Qwen3.5-0.8B, and Liquid AI's LFM2.5-1.2B-Thinking. OpenBMB's own capability benchmarks evaluate all four models across various categories, including general knowledge, domain expertise, coding, instruction adherence, mathematical reasoning, logical reasoning, and agentic tasks. MiniCPM5-1B outperforms its peers in all seven categories, notably excelling in agentic performance and general knowledge.
Evaluation Summary
We conducted three brief assessments. The first involved a classic logic puzzle: "Please act as an expert lawyer and legislator. Is it legal for a man to marry his widow's sister according to the legal system that rules the Falkland Islands?"
The correct response is straightforward—a man with a widow is deceased, and deceased individuals cannot sign marriage certificates. MiniCPM5-1B provided an elaborate explanation of Falkland Islands marriage law but failed to recognize the trap, addressing it instead as a simple jurisdictional issue.
“It is essential to determine the actual marital status in the Falkland Islands, which is a fact to be established by local authorities or through legal processes,” was the model's lengthy response.
Our second evaluation sought a clear A/B choice, but the model opted for neither, providing a both-sides response—a common failure mode observed in smaller models under conversational stress. MiniCPM5-1B is no exception to this trend.
When asked which sector would dominate the economy by the year 2100: Crypto or AI? The model did not engage with the question directly, instead beginning to analyze both cryptocurrency and AI investments as interrelated from the ground up.
None of these results are particularly surprising given the model's 1 billion parameters.
However, the agentic features are the focal point of interest. When paired with an MCP server for web research, MiniCPM5-1B's tendency to generate inaccurate responses to obscure factual queries diminishes significantly.
We asked the model for the current price of bitcoin and for three stock recommendations, and the tool functioned effectively, suggesting Amazon, Microsoft, and Nvidia as viable options.
Final Thoughts
A conversational, locally-deployable agent that can utilize tools, maintain a context of 128K, and operate entirely on-device presents a more intriguing option than merely a standalone Q&A model competing with GPT-4.
However, it’s important not to cancel your AI subscription just yet. Understand its limitations: it has less knowledge compared to larger models, will perform poorly in coding tasks (again, in comparison to larger models), and is far from achieving AGI, if that is your goal.
MiniCPM5-1B is currently available on Hugging Face under an Apache 2.0 license, compatible with vLLM, SGLang, and standard Transformers inference.