# Sam Witteveen — 'MiniCPM5-2B: The Best Sub-Agent Model Yet?'

> Sam Witteveen asks whether MiniCPM5-2B is small enough to run anywhere and still good enough to drive sub-agents.

- Org: Sam Witteveen
- Date: 2026-09-10
- Category: video
- Source: https://www.youtube.com/watch?v=CXvncR7v66o
- Canonical page: https://ai-tldr.dev/releases/sam-witteveen-minicpm5-2b-sep10/

## In short

Sam Witteveen tests MiniCPM5-2B, OpenBMB's small open model, in an episode posted on 10 September 2026. The walkthrough covers the benchmarks, the training recipe including JustRL II, the model variants and a live demo.

## What it is

'MiniCPM5-2B: The Best Sub-Agent Model Yet?' went up on Sam Witteveen's channel on 10 September 2026. The episode looks at MiniCPM5-2B, the small open model OpenBMB published on Hugging Face this month, and asks whether something this size can be trusted with the sub-agent jobs normally handed to a much larger model.

## How it works

The episode is a walkthrough, not a benchmark run. Its own chapter list moves from the benchmarks at 1:06 to the training recipe, spends roughly two and a half minutes on JustRL II, then goes through the MiniCPM5-2B variants and their Hugging Face pages before ending on a live demo at 7:22.

## Why it matters

Sub-agents are where the bill adds up. A coding or research agent can spawn dozens of small calls in one run, and paying frontier prices for every one of them is what makes agent runs expensive. Testing whether a 2B model holds up in that slot is a cost question, not a curiosity.

## Who it's for

developers building multi-agent systems

## Links

- [Watch on YouTube](https://www.youtube.com/watch?v=CXvncR7v66o)
- [Sam Witteveen](https://www.youtube.com/channel/UC55ODQSvARtgSyc8ThfiepQ)
- [Model Card](https://huggingface.co/openbmb/MiniCPM5-2B)
- [GitHub Repo](https://github.com/OpenBMB/MiniCPM)