On-device LLMs gain another contender. OpenBMB, together with the Tsinghua University NLP lab and ModelBest, has released MiniCPM5-2B, the second model in the MiniCPM5 series: a dense 2B Transformer with 2,516,756,480 total parameters (about 1.98B non-embedding), 42 layers, GQA attention with 16 query heads and 2 KV heads, and a native 131,072-token context window, built for phones, laptops and other resource-constrained local deployment. Weights are open under Apache-2.0.

2B versus 4B: how to read the numbers

Across 34 benchmarks the official average is 53.9, while the best 4B-class model in the comparison set, Qwen3.5-4B, sits at 51.1 — the 2B model outscored every larger model in the set. The per-task gaps are wider: 86.5 on AIME 2025 (Qwen3.5-4B: 78.8), 69.1 on LiveCodeBench v6 (56.4), 68.1 on NoLiMa for long context, and 43.5 on BrowseComp-ZH. The standout is 46.4 on SWE-bench Verified — above every 4B-class model compared (LFM2.5-2.6B managed 6.0). These are vendor-reported numbers within the official comparison set, and some scores carry the official Artificial Analysis release mark.

RL teachers plus on-policy distillation

Training runs in three stages: base training, mid-training, then post-training. Post-training starts with 400B tokens of deep-thinking SFT, then trains specialized RL teachers for math, code, agentic tasks and writing, and finally uses On-Policy Distillation (OPD) to distill 16 RL experts, including 5 agentic experts, into one release model. OpenBMB reports the RL+OPD stage added an average 10.96 points to reasoning and general capability, and 6.96 to agentic capability. The data is open too: UltraX web pre-training data, tiered UltraData-Code, 500K agent SFT samples and 80K+ RL samples are all released.

A wide deployment footprint

The model uses the standard LlamaForCausalLM architecture, so mainstream inference engines load it directly — no custom kernels, no model-code fork. Weights ship in GGUF, MLX, GPTQ and DSpark draft-model formats, with official cookbooks for llama.cpp, Ollama, LM Studio, vLLM, SGLang and ArcLight. SGLang provided day-0 support and reported over 250 tokens per second per user on an RTX 5090 with DSpark speculative decoding. On the chip side, the FlagOS platform adapted the model to 9 AI chips including Nvidia, Ascend, Moore Threads and Kunlunxin.

So what

Two pieces of context matter. First, Artificial Analysis released Intelligence Index v4.2 on September 4, three days before the model, with heavier weight on private test sets; MiniCPM5-2B now leads that index among open models under 4B parameters. Second, cryptobriefing notes the 1B model scored 17.9 on the old index while the 2B version lands on the new one — the methodology changed, so cross-version scores are not comparable. For developers, the real signal sits in task-level scores like SWE-bench: a 2B model reaching territory that 4B-class models could not a year ago opens a practical window for on-device agents. But benchmark leads and real-world utility do not always move together — whether it is worth switching is best answered by running it on your own device.

Reference: the official model card (huggingface.co/openbmb/MiniCPM5-2B), plus coverage from cryptobriefing and SGLang.