---
title: "MiniCPM5-2B as a Local Voice Agent | Micdrop"
description: "MiniCPM5-2B tested as the agent of a local call, twice as fast per turn as Qwen3 4B but English only in practice."
url: "https://micdrop.dev/docs/local-models/explorations/minicpm"
---

# MiniCPM5-2B as a Local Voice Agent

Tested as the agent, the part that writes the answers and decides when to call a tool. It is the only model in this section that is recommended, for an English call on a machine with little memory.

MiniCPM5-2B is a 2 billion parameter model from OpenBMB, published as a GGUF, so Ollama pulls it from Hugging Face with `ollama pull hf.co/openbmb/MiniCPM5-2B-GGUF:Q4_K_M`. Like the other agents in this section, it was compared with the recommended model, [Qwen3 4B Instruct](/docs/local-models/agent), using `ollama-agent.ts` with its `OLLAMA_MODEL` and `OLLAMA_LANG` variables, on the same machine and with the same prompt:

The test

MiniCPM5-2B

Qwen3 4B, recommended

Weights in `Q4_K_M`

1.6 GB

2.5 GB

First token, warm, tools declared

~60 ms

~100 ms

A whole turn in English

340 to 820 ms

820 to 1420 ms

Asking the time, with a `get_time` tool

Tool called

Tool called

A meaningless `"uh"`, with `ignore_user_noise`

Answered anyway, 1 call in 9 runs

`CancelLastUserMessage` every run

`"Thanks, goodbye!"`, with `end_call`

No call, offered more help

No call, said goodbye

Numbers written in full, as the prompt asks

`"9:14:07 AM"`

`"nine fifteen forty-three in the morning"`

The model reasons before answering unless reasoning is turned off. Ollama maps `providerOptions: { openai: { reasoningEffort: 'none' } }` to its own `think` flag. A short greeting takes 700 to 1500 ms with reasoning on, and 180 ms with it off.

The model answers a plain `"uh"` instead of calling `ignore_user_noise`: it called the tool once in 9 English runs, while Qwen3 4B called it every time. Noise filtering needs a fallback. It also writes the time as `"9:14:07 AM"` although the prompt asks for numbers in full.

In French it writes `"Il est currently neuf heures"`, invents words such as `"aujourd'nous"` and `"n'hélas"`, and opens every turn with `"Bonjour"`. When the prompt asks for French, the first line of the default Micdrop prompt, _“If you’re first to speak, greet the user and ask how you can help”_, also stops its tool calls. Over 5 identical turns asking the time, `get_time` was called 0 times with that line in the prompt and 5 times with the line removed. The same prompt in English calls the tool 5 times either way. The weights are trained on English and Chinese.

Use MiniCPM5-2B for an English call on a machine with little memory. It is twice as fast per turn as Qwen3 4B and uses 1 GB less. Qwen3 4B Instruct stays the recommended model, because it handles the automatic prompts and other languages. For French tool calls on a machine with more memory, OpenBMB’s larger [MiniCPM-V 4.7 35B-A3B](/docs/local-models/explorations/minicpm-v) handles every automatic tool.
