MiniCPM5-2B

Medium's independent analysis notes that OpenBMB released MiniCPM5-2B on September 7, 2026: a dense, 2.516-billion-parameter causal language model under Apache 2.0 with a native 131,072-token context window. The headline claim is that this 2B-class model is state-of-the-art among sub-4B open models, averaging 53.9 acro The article reports that when MiniCPM5-2B is wired into a real multi-turn agentic harness, its performance collapses by up to 3x relative to single-shot index scores, highlighting a gap between elite single-shot coding benchmarks and chaotic production software engineering realities. It frames the model as benchmark-sp

SitePoint's independent benchmark analysis reports that on Artificial Analysis's GDPval-AA v2, which scores models on real-world work tasks against a 1,000 human baseline, MiniCPM5-2B scores 831 and finishes first among open-weight models under 4B parameters, ahead of Ling 3.0 Tiny (718), Granite 4.2 8B (648), Gemma 4 On the composite Intelligence Index that GDPval-AA v2 feeds into, MiniCPM5-2B places third, behind substantially larger models including Ling 3.0 Tiny at roughly three times the parameters and Gemma 4 12B at several times larger. The article frames the dual result as evidence about how to read benchmark scores and capa

DataNorth reports that OpenBMB released MiniCPM5-2B on 7 September 2026, an open dense model with 2.52 billion parameters under Apache 2.0 and a context window of 131,072 tokens. On OpenBMB's own 34-benchmark comparison table, MiniCPM5-2B averages 53.9 versus 51.1 for Qwen3.5-4B, which carries roughly twice the paramet The model targets hardware a developer controls (laptop, phone, single small GPU) rather than a rented data center, and uses the standard LlamaForCausalLM layout so vLLM, SGLang, Transformers, llama.cpp, Ollama, LM Studio, and MLX load it without custom code. OpenBMB shipped eight artifacts at once: the main model, a b

The vLLM Recipes page documents MiniCPM5-2B as the 2B dense checkpoint in OpenBMB's MiniCPM5 series, built for on-device and resource-constrained deployment with strong performance on agentic tool use, code generation, and reasoning. It uses the standard LlamaForCausalLM architecture, so vLLM loads it natively with no At 2B parameters the BF16 weights (5 GB) fit on a single consumer GPU at TP=1, and the model supports the full native 128K context window, which can be reduced via --max-model-len to free KV cache on smaller GPUs. The page lists supported serving targets including AMD MI300X (192G), MI325X (256G), MI355X (288G), and MI