MiniCPM 5-2B: LM Studio setup and agent tests

MiniCPM 5-2B is tested as a local LLM through LM Studio's API, with tool calls and automated code checks. The speaker uses the official Q4KM GGUF, reports a 1.56 GB download, and configures an 8K context window with GPU acceleration disabled. An initial runtime failure clears after switching to a CPU AVX2 runtime. The reported load uses approximately 1.45 GB of memory.

The 12 tests examine practical AI agent behavior. The model selects a calculator, completes a tool execution loop, preserves strict task arguments, and prepares two independent calls in one response. That last result demonstrates parallel planning, rather than simultaneous service execution. A dependent customer lookup initially fails because the model supplies a placeholder ID; explicit instructions to wait resolve it.

The speaker reports successful timeout retry handling and resistance to destructive tool distractors and a tool-output prompt injection. Missing information causes a stall: the model notices an absent due date but exhausts its 512-token output budget without asking for it.

For coding, a Python task manager initially passes four of five tests. After receiving the failure, the model fixes its summary format and passes all five. A subsequent multi-file task produces no file-writing calls. The final tally is eight clean passes, two conditional or partial results, and two failures. The speaker recommends bounded workflows and a coding assistant backed by automated validation, rather than fully autonomous coding.