Summary
Sam Witteveen reviews MiniCPM5, a 1-billion-parameter dense model from OpenBMB (Tsinghua NLP Lab), framing it against Andrej Karpathy’s “cognitive core” concept — a tiny model that strips encyclopedic knowledge in favor of reasoning and tool use, small enough to run on smartphones several years old. The video situates MiniCPM5 in a growing class of sub-2B models alongside TinyLlama, Meta’s Llama 3.2 1B, Qwen 0.5B-3B, Gemma 4, and LFM variants, noting that MiniCPM5 beats Qwen 3 0.5B on several benchmarks despite being a non-reasoning model.
Witteveen pays particular attention to the model’s AA Omniscience benchmark score (near zero, meaning it hallucinates rarely when it doesn’t know an answer) and its strong function-calling and tool-use capabilities relative to its size — properties he argues are more important than raw accuracy for agentic applications. The architecture is Llama-style, enabling easy integration into existing tooling.
The practical portion examines two “mini harnesses” built around the model: a Rust-based edge smart home controller developed by a third-party developer, and OpenBMB’s own MiniCPM Desk Pet, a full Electron app that runs the model locally. Witteveen’s broader argument is that 1B models are increasingly viable for embedding a layer of intelligence into hardware and apps that previously had none — driven in part by phone manufacturers deploying on-device models with swappable LoRA fine-tunes for specific use cases.
📺 Source: Sam Witteveen · Published July 05, 2026
🏷️ Format: Review
