OpenBMB：清华大学开源社区如何推动大语言模型高效计算与参数微调

Introduction

OpenBMB (Open Benchmark for Language Models) is an open-source community and toolkit ecosystem initiated by a team from Tsinghua University since 2018. It is dedicated to constructing a comprehensive and efficient computational toolchain for the entire lifecycle of large-scale pre-trained language models (LLMs). A core contribution of the project lies in pioneering research on parameter-efficient fine-tuning (PEFT) methods. The community has garnered significant global recognition, accumulating over 4,000 stars on GitHub and receiving the Best Demo Paper Award at ACL 2022, a top-tier conference in natural language processing. The toolkit includes components like OpenDeltaOpenBMB工具库中的组件，支持研究者在各类预训练模型中实现增量微调方法。, enabling researchers to implement incremental fine-tuning methods across various pre-trained models [1].

OpenBMB（大型语言模型开放基准）是一个始于2018年、由清华大学团队支持发起的开源社区与工具包生态系统。其致力于构建覆盖大规模预训练语言模型全流程的高效计算工具体系。该项目的核心贡献之一在于对参数高效微调只训练大模型中少量参数的技术，如LoRA和QLoRA，大幅降低微调所需的计算资源和显存消耗。（PEFT）方法的开创性研究。该社区获得了广泛的全球关注，在GitHub上累计获得超过4000星标，并获得了自然语言处理领域顶级国际会议ACL 2022的最佳系统演示论文奖。其工具库包含OpenDeltaOpenBMB工具库中的组件，支持研究者在各类预训练模型中实现增量微调方法。等组件，支持研究者在各类预训练模型中实现增量微调方法 [1]。

Core Philosophy and Historical Context

Since 2018, the Tsinghua team has been consistently engaged in innovative research on large language models while steadfastly building the OpenBMB open-source community [1]. The philosophy centers on lowering the barriers to cutting-edge LLM research and application by providing robust, standardized tools. This long-term commitment has fostered a collaborative environment where academic advancements are rapidly translated into accessible resources for the wider AI community.

自2018年以来，清华大学团队始终坚持开展大语言模型的创新研究，并持续建设OpenBMB开源社区 [1]。其核心理念是通过提供强大、标准化的工具，降低前沿大模型研究和应用的门槛。这一长期承诺培育了一个协作环境，使得学术进展能够快速转化为更广泛AI社区可获取的资源。

Key Tools and Frameworks

OpenDeltaOpenBMB工具库中的组件，支持研究者在各类预训练模型中实现增量微调方法。 and Parameter-Efficient Fine-Tuning

A foundational component of OpenBMB is its focus on parameter-efficient fine-tuning. Traditional fine-tuning of massive LLMs requires updating all parameters, which is computationally expensive. OpenBMB's methods, facilitated by tools like OpenDeltaOpenBMB工具库中的组件，支持研究者在各类预训练模型中实现增量微调方法。, allow for selective updates or the introduction of small, trainable adapter modules into a frozen pre-trained model. This dramatically reduces computational cost and storage requirements while maintaining high performance on downstream tasks.

OpenBMB的一个基础组件是其对参数高效微调只训练大模型中少量参数的技术，如LoRA和QLoRA，大幅降低微调所需的计算资源和显存消耗。的关注。传统的大模型微调需要更新所有参数，计算成本高昂。OpenBMB提供的方法，通过OpenDeltaOpenBMB工具库中的组件，支持研究者在各类预训练模型中实现增量微调方法。等工具实现，允许对冻结的预训练模型进行选择性更新或引入小型可训练的适配器模块。这极大地降低了计算成本和存储需求，同时在下游任务上保持高性能。

UltraEval-Audio由清华大学NLP实验室、OpenBMB与面壁智能联合开源的音频模型测评框架，支持音频模型的一键测评与复现。: Benchmarking Audio Models

On January 4, 2026, the Tsinghua NLP Lab, OpenBMB, and ModelBest AI jointly released and open-sourced UltraEval-Audio由清华大学NLP实验室、OpenBMB与面壁智能联合开源的音频模型测评框架，支持音频模型的一键测评与复现。, an evaluation framework for audio models. Its v1.1.0 version significantly enhanced usability by adding one-click reproducibility for popular audio models and extended support for specialized models like Text-to-Speech (TTS), Automatic Speech Recognition (ASR), and Codec. The version introduced an isolated inference mechanism to lower the barrier for model reproduction. UltraEval-Audio由清华大学NLP实验室、OpenBMB与面壁智能联合开源的音频模型测评框架，支持音频模型的一键测评与复现。 v1.1.0 has become a crucial evaluation tool for many influential audio and multimodal models, such as MiniCPM-o2.6 and VoxCPM. The framework's open-source release aims to standardize audio model evaluation and propel research forward [3].

2026年1月4日，清华大学NLP实验室、OpenBMB与面壁智能联合发布并开源了UltraEval-Audio由清华大学NLP实验室、OpenBMB与面壁智能联合开源的音频模型测评框架，支持音频模型的一键测评与复现。，一个面向音频模型的测评框架。其v1.1.0版本通过为热门音频模型增加一键复现能力，并扩展对文本转语音（TTS）、自动语音识别（ASR）、编解码（Codec）等专业模型的支持，显著提升了易用性。该版本引入了隔离推理的运行机制，旨在降低模型复现的门槛。UltraEval-Audio由清华大学NLP实验室、OpenBMB与面壁智能联合开源的音频模型测评框架，支持音频模型的一键测评与复现。 v1.1.0已成为MiniCPM-o2.6、VoxCPM等众多高影响力音频及全模态模型的重要测评工具。该框架的开源旨在推动音频模型评测的标准化与研究进展 [3]。

AgentCPM-Explore由清华大学NLP实验室、中国人民大学、面壁智能及OpenBMB开源社区共同发布的40亿参数智能体模型，专注于智能体任务，具有高参数效率。: A Compact Yet Powerful Agent Model

On January 14, 2026, the Tsinghua NLP Lab, in collaboration with Renmin University of China, ModelBest AI, and the OpenBMB community, released AgentCPM-Explore由清华大学NLP实验室、中国人民大学、面壁智能及OpenBMB开源社区共同发布的40亿参数智能体模型，专注于智能体任务，具有高参数效率。, a 4-billion-parameter agent model with potential for on-device deployment. It demonstrated exceptional parameter efficiency across multiple agent evaluation benchmarks, including GAIA, HLE, Browsercomp, and their Chinese variants. Its performance is comparable to or surpasses that of advanced models with 8 billion parameters, and in some aspects rivals models with over 30 billion parameters or even closed-source LLMs. Notably, on the Xbench-DeepResearch evaluation, it outperformed closed-source systems like OpenAI-o3 and Claude-4.5-Sonnet. The model has been fully open-sourced, along with its complete training and evolution code pipeline [4].

2026年1月14日，清华大学自然语言处理实验室联合中国人民大学、面壁智能及OpenBMB开源社区共同发布了AgentCPM-Explore由清华大学NLP实验室、中国人民大学、面壁智能及OpenBMB开源社区共同发布的40亿参数智能体模型，专注于智能体任务，具有高参数效率。，这是一个参数规模为40亿、具备终端设备部署潜力的智能体模型。在GAIA、HLE、Browsercomp及其中文版本等多项智能体评测基准测试中，它展现出了优异的参数效率。其性能接近或超越了参数量达80亿的先进模型，部分表现可与300亿以上参数乃至闭源大模型相媲美。在Xbench-DeepResearch测评中，该模型的表现优于OpenAI-o3、Claude-4.5-Sonnet等闭源系统。该模型已全面开源，同步公开了完整的训练与演进代码流程 [4]。

Impact and Future Direction

OpenBMB has established itself as a significant force in democratizing LLM technology. By providing essential tools for efficient fine-tuning, comprehensive evaluation (as seen with UltraEval-Audio由清华大学NLP实验室、OpenBMB与面壁智能联合开源的音频模型测评框架，支持音频模型的一键测评与复现。), and compact yet capable model architectures (exemplified by AgentCPM-Explore由清华大学NLP实验室、中国人民大学、面壁智能及OpenBMB开源社区共同发布的40亿参数智能体模型，专注于智能体任务，具有高参数效率。), the project directly addresses key challenges in the field: computational cost, standardization, and practical deployment. The community's active development and high-impact releases suggest a continued focus on bridging the gap between state-of-the-art research and accessible, practical implementation. Future contributions will likely further expand into multimodal evaluation, agent system tooling, and even more efficient model architectures, solidifying its role as a cornerstone of the open-source LLM ecosystem.

OpenBMB已成为推动大模型技术平民化的一支重要力量。通过提供参数高效微调只训练大模型中少量参数的技术，如LoRA和QLoRA，大幅降低微调所需的计算资源和显存消耗。、全面测评（如UltraEval-Audio由清华大学NLP实验室、OpenBMB与面壁智能联合开源的音频模型测评框架，支持音频模型的一键测评与复现。）以及紧凑而强大的模型架构（如AgentCPM-Explore由清华大学NLP实验室、中国人民大学、面壁智能及OpenBMB开源社区共同发布的40亿参数智能体模型，专注于智能体任务，具有高参数效率。）等关键工具，该项目直接应对了该领域的核心挑战：计算成本、标准化和实际部署。该社区的活跃开发和高影响力发布表明，其将继续致力于弥合前沿研究与可访问、可实践的应用之间的鸿沟。未来的贡献可能会进一步扩展到多模态评测、智能体系统工具以及更高效的模型架构，巩固其作为开源大模型生态系统基石的定位。

References

[1] OpenBMB Community & Tools Overview
[2] UltraRAG 2.1OpenBMB联合发布的技术或工具版本，具体功能未在内容中详细说明，但属于OpenBMB生态中的重要发布。 Release
[3] UltraEval-Audio由清华大学NLP实验室、OpenBMB与面壁智能联合开源的音频模型测评框架，支持音频模型的一键测评与复现。 v1.1.0 Release Announcement
[4] AgentCPM-Explore由清华大学NLP实验室、中国人民大学、面壁智能及OpenBMB开源社区共同发布的40亿参数智能体模型，专注于智能体任务，具有高参数效率。 Release Announcement