MiniCPM5-1B review: tool use, desk pet and GLM 5.2

MiniCPM5-1B is the focus of this review of small models for on-device applications. The speaker describes OpenBMB's text-only model as a 1B dense model with a Llama-style architecture, a 128K context window and an Apache 2 license. The discussion covers its base, supervised fine-tuned and fully trained releases, along with published training datasets.

The practical question is how well this local LLM handles tool use. In the speaker's tests, it calls weather and calculator functions with suitable arguments, handles repeated lookups and completes basic multistep tasks. Longer AI agent workflows are less reliable: one run completes about 12 tool calls, while another fails. A comparison with GLM 5.2 shows more consistent long workflows in those tests, though the speaker notes the size mismatch and runs GLM remotely.

A MiniCPM Desk Pet demonstration shows an Electron app running a downloaded GGUF model, with chat controls and selectable LoRA variants. The speaker questions its practical usefulness and flags its license as a limitation without specifying the terms.

General chat and generation tests expose other weaknesses. The model struggles to adopt an assigned name, misses a requested 5,000-word essay and produces a poor bicycle SVG. In GSM8K and MMLU evaluations, the speaker attributes many failures to repetitive reasoning or exhausted token budgets, including runs capped at 16,000 tokens.