MiniCPM5-2B vs Qwen3.5-4B: local setup and coding tests

MiniCPM5-2B is a local LLM aimed at coding and tool use. The speaker describes the text-only model as a dense 2.5-billion-parameter network with 42 layers, roughly 128,000 tokens of native context, and an Apache 2.0 license. Its Llama architecture replaces MiniCPM4's custom attention approach.

The walkthrough uses MLX on a Mac for generation and interactive chat, then connects an OpenAI client to llama.cpp for a weather function call. The MLX demonstration uses a 1.3 GB file. The speaker warns that older MLX-LM versions ignore the model's end-of-sequence list.

The reported vendor benchmarks give MiniCPM5-2B higher scores than Qwen3.5-4B on LiveCodeBench and SWE-bench Verified, but lower scores on SWE-bench Pro and Terminal Bench. These results support a narrower coding comparison, rather than a general claim of equal capability.

Sampling settings are a practical concern. The speaker cites a GitHub report where Q4_K_M generations repeated uncontrollably 92% of the time. A repeat penalty of 1.15 improved the reported scores; OpenBMB later advised setting min_p to zero.

The speaker also reports manual GGUF setup for LM Studio and a model file requirement for Ollama. Thinking is enabled by default, and the documented sampling profile covers that mode. English and Chinese are the stated languages. For larger tasks, the speaker still prefers Qwen3.5-4B.