> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# OpenBMB Ships MiniCPM5-2B, a 2.52B Dense Model Aimed at On-Device Agents
- URL: https://globalfeed.ai/en/openbmb-ships-minicpm5-2b-a-2-52b-dense-model-aimed-at-on-device-agents/
- Published: 2026-09-08T09:50:07.000Z
- Updated: 2026-09-08T09:50:07.000Z
- Description: OpenBMB's Apache 2.0 licensed MiniCPM5-2B reports a 53.9 average across 34 benchmark rows, with its widest leads in tool calling and coding agents.
- Author: GlobalFeed Editor
- Tags: OpenBMB, MiniCPM5, open source models, x-OpenBMB, dil-en, denetim-gerekli, video, x-gitti

OpenBMB has released **MiniCPM5-2B**, the second checkpoint in the MiniCPM5 series after MiniCPM5-1B. It is a dense causal language model with 2,516,756,480 parameters, 1,981,982,720 of them outside the embeddings. The model uses 42 layers, grouped-query attention with 16 query heads and 2 key/value heads, and a native context window of 131,072 tokens. Because the architecture is standard LlamaForCausalLM, mainstream engines load it without custom kernels or a model-code fork.

The weights are Apache 2.0 and run on vLLM, SGLang, Transformers, llama.cpp, Ollama, LM Studio, MLX and FlagOS.

## Benchmark numbers

OpenBMB compares the model with LFM2.5-2.6B, Qwen3.5-2B and Gemma-4-E2B-it in the same size class, and lists Qwen3.5-4B, granite-4.2-3B, Nemotron-3-Nano-4B, Gemma-4-E4B-it and LFM2.5-8B-A1B for reference. Across 34 benchmark rows it averages 53.9\. The strongest baseline in that set is Qwen3.5-4B at 51.1, followed by granite-4.2-3B at 42.7 and LFM2.5-2.6B at 33.2.

On code reasoning it posts 69.1 on LiveCodeBench v6 against 56.4, and 46.4 on SWE-bench Verified against 33.6\. Tool use shows the widest margin: 97.1 on τ2-Bench Telecom, 66.6 on BFCL v4 and 20.8 on τ3-Bench Banking against 6.8\. Long context is mixed: 68.1 on NoLiMa against 43.5, but 59.0 on AA-LCR against 61.0 and 43.7 on LongBench v2 against 47.3\. General knowledge is where the size gap shows, with 70.8 on MMLU-Pro against 78.0 and 8.9 on Humanity's Last Exam against 9.9\. Rows sourced from Artificial Analysis are marked separately from internally reproduced ones.

## Training and released data

Training follows the UltraData tiered data management method. Post-training begins with 400B tokens of deep-thinking SFT, then trains separate RL teachers for math, code, agentic tasks and writing using the critic-based JustRL II algorithm. The final step, on-policy distillation, merges 16 RL experts, five of them agentic, into one shipped model, computing full-vocabulary reverse KL divergence between student and teacher logits as the advantage estimate. OpenBMB measures the RL plus OPD stage at 10.96 average points on reasoning and general benchmarks and 6.96 points on agentic ones.

Alongside the weights, OpenBMB published Ultra-FineWeb, Ultra-FineWeb-L3, UltraX, UltraData-Code, UltraData-Math, UltraData-SFT-2605, UltraData-SFT-Agent-2609 with 500K agent samples and UltraData-RL-2609 with more than 80K RL samples, plus Base, Midtrain and SFT-only intermediate checkpoints. Full details are in the [MarkTechPost report](https://www.marktechpost.com/2026/09/07/openbmb-releases-minicpm5-2b-a-2-52b-dense-model-averaging-53-9-across-34-benchmarks-and-built-to-run-on-device/?ref=globalfeed.ai).