Summary
Sam Witteveen examines MiniCPM 2B (technically 2.5B), the latest small model from OpenBMB, which claims to beat 4B-class models on function calling and agentic tasks. The video walks through the model’s updated training recipe: supervised fine-tuning on 400 billion deep-thinking tokens (up from 200B for the 1B predecessor), followed by a new reinforcement learning stage called RL2 that incorporates a critic model for improved training stability, and a final on-policy distillation layer where the model refines its own trajectories against teacher model outputs.
Benchmark results are presented honestly. MiniCPM 2B edges out Qwen 3.5 4B on SWE-bench Verified and GPQA Diamond while running at a fraction of the parameter count, but falls significantly behind Qwen on SWE-bench Pro and Terminal Bench — approximately half the score in those categories. Witteveen attributes much of Qwen’s advantage to a substantially larger pre-training budget.
For the local inference tests, Witteveen runs three variants — BFloat16, a DSpark speculative decoding build, and a 4-bit GGUF via llama.cpp — all on an RTX Pro 6000. The model handles logic puzzles and hard math well, reflecting its RL-heavy training focus, but struggles with long-form prose generation and complex visual tasks like SVG rendering. The overall verdict: a compelling sub-agent or tool-calling component for resource-constrained pipelines, but not a general-purpose replacement for larger models.
📺 Source: Sam Witteveen · Published September 10, 2026
🏷️ Format: Benchmark Test
