MiniCPM5-2B review: local tool use and Qwen comparison

MiniCPM5-2B is assessed as a local LLM for sub-agents that call tools. The speaker reviews OpenBMB's training recipe, including 400 billion deep-thinking tokens for supervised fine-tuning, critic-based JustRL II, and on-policy distillation. He describes benchmark gains over other small models, but notes that Qwen remains substantially ahead on SWE-bench Pro and Terminal Bench.

The local demonstration compares BFloat16 and DeepSpark speculative-decoding variants through SGLang with a 4-bit GGUF version through llama.cpp. All run on an RTX Pro 6000, so the reported speeds do not establish performance on consumer hardware. The speaker reports more than 400 tokens per second with GGUF. His SGLang setup returns incorrect tool-call formats, which he attributes to configuration rather than the model.

The GGUF demonstration passes eight advanced tool-use tests in one run, including planning and recovery from deliberate failures. Repeated runs reveal a limitation: the speaker reports roughly 80% success on the prompt-injection test. He finds logic and math results useful, while the requested 5,000-word essay falls well short and SVG and HTML generation disappoint.

His recommendation is to use the model for an AI agent that delegates to tools rather than relies on internal knowledge. A sponsored RunPod Serverless segment also demonstrates a managed GPU endpoint with an OpenAI-compatible API and configurable scaling.