ModelBest Open-Sources MiniCPM5-2B: No. 1 Under 4B on the AA Leaderboard, with Early General On-Device Agent Capabilities

2026-09-08

On September 8, ModelBest and the OpenBMB open-source community officially released the new-generation "small powerhouse" MiniCPM5-2B.

With 2B parameters, MiniCPM5-2B is a high-density on-device language foundation model that supports tool calling, deep search, code generation, and other tasks, taking an early step toward a general on-device agent.

According to the latest evaluation by the internationally recognized Artificial Analysis Intelligence Index (hereafter "the AA Leaderboard"), MiniCPM5-2B scores 23 points overall, ranking first among open-source foundation models under 4B parameters worldwide.

On Agentic Index, which measures agent capabilities, MiniCPM5-2B leads by a wide margin: it scores 20 points, while peer models such as LFM2.5-2.6B, Granite 4.2 3B, and Mistral 3 3B score only 2.

Behind this leap in capability are ModelBest's continued innovations in underlying large-model technologies — data, model training, and frameworks — as well as further validation of the "Density Law" first proposed by ModelBest on an on-device foundation model.

Best Overall Performance Under 4B Worldwide

On the AA Leaderboard, MiniCPM5-2B raises its overall score to 23 points, placing it first among open-source models under 4B parameters worldwide, ahead of Gemma 4 12B, Qwen3.5 9B, and the major mainstream 3B–4B models:

In terms of intelligence density, MiniCPM5-2B achieves an extreme parameter-efficiency ratio: with only 2B parameters, it leads 4B peer open-source models of twice the size in intelligence.

This means: MiniCPM5-2B uses roughly half the parameters to deliver nearly double the intelligence score, giving it an extremely high intelligence density per parameter.

In inference efficiency and output quality, MiniCPM5-2B shows extremely high token efficiency:

Consuming the same 21k tokens (14k thinking + 7k answer), MiniCPM5-2B scores 23 points, first in its class. By comparison, Granite 4.2 3B consumes 19k tokens (12k thinking + 7k answer) and scores only 14 points; LFM2.5-2.6B consumes 21k tokens (14k thinking + 7k answer) and scores only 11 points.

In other words, MiniCPM5-2B reaches nearly twice the intelligence at equal or even tighter compute cost:

In general agent capabilities, MiniCPM5-2B also demonstrates the potential of a high-density on-device language foundation model:

According to the AA Leaderboard, MiniCPM5-2B scores a commanding 20 points, a 10x capability gain over peer models (including LFM2.5-2.6B, Granite 4.2 3B, and Mistral 3 3B), an early form of high-intelligence-density general on-device agent capability.

Beyond the AA Leaderboard, MiniCPM5-2B also performs strongly on 34 benchmarks spanning code reasoning, math reasoning, instruction following, general knowledge, long text, tool calling, and agent tasks:

Its average score reaches 53.9, ranking first — ahead of the second-place 2B model at 33.2 and above the best-performing 4B model, Qwen3.5-4B.

On specific tasks, it leads peer 2B models and even larger mainstream 3B–4B industry models in code reasoning, math reasoning, long text, tool calling, search agents, and general agents.

With 1,000 points as the human baseline, MiniCPM5-2B scores 891 on real-world task evaluations, topping models under 4B and far exceeding Gemma 4 12B (647 points), which has about six times its parameters, bringing it closer to human level.

MiniCPM5-2B's performance in intelligence density and token consumption further validates the scientific basis of the "Density Law" first proposed by ModelBest, and demonstrates its advantage in balancing performance and cost on on-device devices with limited compute and VRAM, such as phones, car systems, PCs, and robots.

Data Governance for More Efficient Intelligence

Behind MiniCPM5-2B's efficient balance of performance and cost is squeezing every last bit of potential out of every model parameter. What supports this result is ModelBest's data governance experience and achievements accumulated since 2023.

Limited by compute, VRAM, and other factors, the parameter scale of models that can currently be deployed on edge devices is limited, so the intelligence density of every parameter directly determines the ceiling of on-device intelligence. Data governance is the key to improving intelligence density.

MiniCPM5-2B builds on ModelBest's UltraData L0–L4 tiered data governance system and proposes comprehensive data governance schemes across general web, math, code, SFT, RL, and other dimensions. Alongside MiniCPM5-2B, four datasets were also newly open-sourced: UltraData-Code, UltraData-SFT-Agent-2609, UltraData-RL-2609, and UltraX.

UltraData-Code:

Aimed at code pre-training and capability optimization, covering UltraData L0–L3 tiered code data. The project collects raw data from about 192 million public GitHub repositories and, through cleaning and deduplication, algorithmic code filtering, and task-oriented structured synthesis, produces about 400B tokens of L2 finely filtered algorithmic code data and about 150B tokens of L3 task-oriented synthetic data.

UltraData-SFT-Agent-2609:

Used for agent capability optimization during MiniCPM5-2B's post-training, containing about 500,000 training samples covering tool calling, web search, code generation, Office document processing, database interaction, and other tasks, effectively improving the agent capabilities of on-device models.

UltraData-RL-2609:

This is the core dataset for MiniCPM5-2B's reinforcement learning stage, containing more than 80,000 high-quality samples covering math reasoning, code generation, general knowledge Q&A, and long-text understanding. To address problems such as unverifiable questions, untrustworthy reward labels, and difficulty mismatched to model capability, the dataset is strictly cleaned through a three-stage pipeline of verifiability filtering, reward-trustworthiness validation, and difficulty matching, providing high signal-to-noise reward signals and a reliable data foundation for RL training.

UltraX:

UltraX is a function-calling-style refinement framework for large-scale pre-training data. It uses a lightweight model to generate a processing script for each data item, enabling low-cost, fine-grained, and traceable data cleaning. UltraX contains five data subsets totaling about 100B tokens and 114 million samples, effectively improving pre-training data quality.

UltraData datasets have been widely validated in the open-source community. As of September 2026, UltraData series datasets have been downloaded more than 2.66 million times in total, and more than 10 datasets have been open-sourced.

Scalable Reinforcement Learning

With the open-sourcing of MiniCPM5-2B, ModelBest and OpenBMB also open-sourced the new self-developed reinforcement learning framework Meshy and the reward-based reinforcement learning strategy JustRL II.

On the training framework side, Meshy completely removes the central controller and Ray dependency from RL training, modeling training, inference, reward computation, and all other roles as peer services, and uses data-readiness states in TransferQueue to drive the actual control flow. It can easily switch among synchronous, asynchronous, and fully asynchronous training modes, making it easy to extend.

At the algorithm level, when on-device models do long chain-of-thought reinforcement learning, scores often fall instead of rising during training: on one hand, training data is noisy and difficulty-mismatched, so reward signals easily mislead the model; on the other hand, GRPO's credit assignment is too coarse — faced with reasoning chains of tens of thousands of tokens, it cannot tell which step is right and which is wrong. To address this, the MiniCPM team proposes JustRL II: for data, a three-stage pipeline filters and calibrates training data; for the algorithm, GRPO is equipped with a Critic to achieve token-level credit assignment. With JustRL II, MiniCPM5-2B achieves significant performance gains in RL post-training.

Day 0 Adaptation on Multiple Chips to Accelerate Industrial Deployment

To accelerate industrial deployment, MiniCPM5-2B has completed Day 0 native adaptation with Intel, Rockchip, Arm, and other chip or computing platform vendors, performing well on multiple chips.

On the Intel platform, leveraging the full heterogeneous compute (CPU, GPU, NPU) of the Intel Core Ultra series XPU and the OpenVINO™ toolkit, MiniCPM5-2B has undergone deep optimization across the full chain of kernels, graph fusion, and memory scheduling. Compared with the baseline version, prefill and decode performance are both substantially improved; runtime memory usage can be significantly reduced and long-context inference latency greatly cut, running deep-thinking reasoning, tool calling, deep search, and code generation — the full set of on-device agent capabilities — stably at low power. Developers can achieve out-of-the-box local deployment on Intel AI PCs, forming a complete loop from underlying hardware and optimization toolchains to upper-layer agent applications, fully unleashing the industrial potential of MiniCPM5-2B's on-device AI.

On the Rockchip platform, MiniCPM5-2B can achieve complete on-device deployment on the RK3588+RK1828 dual-chip platform. With Rockchip's RKNN3 full toolchain, it completes model quantization and deep kernel optimization, and with the high bandwidth of the RK182X series, it effectively reduces the first-token latency of large models, running stably on on-device hardware and fully reproducing core on-device agent capabilities including deep-thinking reasoning, tool calling, deep search, and code generation.

On the Arm computing platform, MiniCPM5-2B can now run efficiently on mobile devices powered by Armv9 computing platforms with second-generation Scalable Matrix Extension (SME2) technology. Built into the Armv9 CPU architecture, Arm SME2 enables AI acceleration directly on the CPU and efficiently supports various real-time mobile inference tasks under AI heterogeneous computing frameworks. Actual test results show that when running MiniCPM5-2B on mobile devices with SME2 enabled, prefill and decode performance improved by about 1.7x and 1.2x respectively, helping deliver a smoother, smarter on-device AI experience.

Data, Framework, and Training Method All Open-Sourced

In this joint open-source release, ModelBest and OpenBMB open-source not only the MiniCPM5-2B model itself but also the training Recipe, framework, and data behind it.

In recent years, ModelBest and the OpenBMB open-source community have continued to advance the openness of foundation model technology. As of August 2026, MiniCPM series models have surpassed 50 million open-source downloads, and the MiniCPM language foundation model has surpassed 13 million downloads.

MiniCPM5-2B is now integrated into mainstream development toolchains and adapted to a variety of model development and deployment tools.

For model training and fine-tuning, MiniCPM5-2B supports tools such as LlamaFactory and ms-swift; for inference deployment, it supports mainstream frameworks including SGLang, vLLM, llama.cpp, Ollama, and Hugging Face Transformers, and is compatible with ModelBest's self-developed Arclight CPU inference framework. Developers can choose the deployment and development scheme that suits their device type, compute conditions, and application needs.

Advancing China's AGI technology through open source has always been the shared pursuit of ModelBest and OpenBMB. Going forward, ModelBest will continue to uphold the open-source philosophy, accelerating the deployment and accessibility of AI so that stronger on-device intelligence reaches a wider range of real-world scenarios.

Currently, the MiniCPM5-2B model, data, and framework are all open-sourced, and developers can access them at the following addresses.

➤ MiniCPM5-2B open-source links (including the model and UltraData series datasets)

HuggingFace:

🔗 https://huggingface.co/collections/openbmb/minicpm5

GitHub:

🔗 https://github.com/OpenBMB/MiniCPM

➤ Meshy framework open-source links

GitHub:

🔗 https://github.com/OpenBMB/Meshy

Blog:

🔗 https://maydomain.notion.site/meshy-blog-zh

➤ JustRL II reinforcement learning algorithm

Blog:

🔗 https://panhaoxuan.notion.site/justrl-ii-small-llms-to-128k-reasoning-with-a-critic-cn