AI News Feed
Market watch
TP-Link launches first Wi-Fi 8 routers at IFA 2026, but experts say most users should waitHackers return most of $320M bitcoin stolen from Liquid Network after patchEthernet splitter vs switch: What's the difference and how do you use them?U.S. Educators Embrace AI in Lessons Instead of Policing ItGrapheneOS delays overhauled Messages app, announces Android fixesOpenBMB Releases MiniCPM5-2B, a 2.52B-Parameter On-Device Model Averaging 53.9 Across 34 Benchmarks
TP-Link launches first Wi-Fi 8 routers at IFA 2026, but experts say most users should waitHackers return most of $320M bitcoin stolen from Liquid Network after patchEthernet splitter vs switch: What's the difference and how do you use them?U.S. Educators Embrace AI in Lessons Instead of Policing ItGrapheneOS delays overhauled Messages app, announces Android fixesOpenBMB Releases MiniCPM5-2B, a 2.52B-Parameter On-Device Model Averaging 53.9 Across 34 Benchmarks
Large Language Models

OpenBMB Releases MiniCPM5-2B, a 2.52B-Parameter On-Device Model Averaging 53.9 Across 34 Benchmarks

OpenBMB released MiniCPM5-2B, a 2.52B-parameter dense on-device model averaging 53.9 across 34 benchmarks.

AI News Feed · 2026-09-08 07:18

The second checkpoint in the MiniCPM5 series and the follow-up to MiniCPM5-1B, MiniCPM5-2B has 2,516,756,480 parameters, 42 layers, grouped-query attention with 16 query heads and two key/value heads, and a native context window of 131,072 tokens. It uses a standard LlamaForCausalLM architecture, requires no custom kernels or model-code forks, and ships under the Apache 2.0 license. The report says it runs through vLLM, SGLang, Transformers, llama.cpp, Ollama, LM Studio, MLX and FlagOS.

According to MarkTechPost, OpenBMB compares MiniCPM5-2B against models in the same size range, including LFM2.5-2.6B, Qwen3.5-2B and Gemma-4-E2B-it, 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, above Qwen3.5-4B at 51.1, granite-4.2-3B at 42.7 and LFM2.5-2.6B at 33.2. On code reasoning it scores 69.1 on LiveCodeBench v6 and 46.4 on SWE-bench Verified. Its widest lead appears in tool use, with 97.1 on τ²-Bench Telecom, 66.6 on BFCL v4 and 20.8 on τ³-Bench Banking. Long-context results are mixed: 68.1 on NoLiMa but 59.0 on AA-LCR and 43.7 on LongBench v2. The model trails on knowledge-heavy tests, posting 70.8 on MMLU-Pro and 8.9 on Humanity's Last Exam. The report notes OpenBMB separately marks rows sourced from Artificial Analysis and internally reproduced ones.

The training recipe runs in stages: SFT, reinforcement learning and then on-policy distillation, following the UltraData tiered data management method. After base training and mid-training, OpenBMB applies 400 billion tokens of deep-thinking SFT, then trains specialized RL teachers for mathematics, code, agentic tasks and writing using the critic-based JustRL II algorithm. The final distillation merges 16 RL experts, five of them agentic, into the released model, computing full-vocabulary reverse KL divergence between student and teacher logits at each response position. The report says the RL plus distillation stage contributed 10.96 average points on reasoning and general benchmarks and 6.96 points on agentic benchmarks.

Alongside the weights, OpenBMB released the datasets used for training: Ultra-FineWeb, Ultra-FineWeb-L3, UltraX, UltraData-Code, UltraData-Math, UltraData-SFT-2605, UltraData-SFT-Agent-2609 with 500,000 agent samples, and UltraData-RL-2609 with more than 80,000 RL samples. Intermediate checkpoints covering base, mid-train and SFT-only stages are also published, allowing each training stage's contribution to be measured.

In MarkTechPost's assessment, MiniCPM5-2B is a credible on-device choice for agentic and tool-calling workloads rather than a general knowledge model. Its advantages are strongest in tool use, coding agents and NoLiMa-style long-context retrieval; it trails larger models on MMLU-Pro, GPQA-Diamond and MATH-500. The open weights, data and intermediate checkpoints make the reported training gains directly checkable.