在本报告中，我们介绍了 Command A 的开发，这是一款专为真实企业场景设计的强大大型语言模型。Command A 是一款面向智能体优化且支持多语言的模型，支持全球商业中使用的 23 种语言，并采用新颖的混合架构，兼顾效率与顶尖性能。它提供了业界领先的检索增强生成（RAG）能力，具备事实锚定和工具使用功能，可自动化复杂的业务流程。这些能力通过去中心化的训练方法实现，包括自优化算法和模型合并技术。我们还提供了 Command R7B 的结果，该模型与 Command A 在能力和架构上具有相似性。两个模型的权重均已发布以供研究使用。本技术报告详细介绍了我们的原创训练流程，并在一系列企业相关任务和公开基准上对我们的模型进行了广泛评估，展示了出色的性能和效率。
In this report we describe the development of Command A, a powerful large language model purpose-built to excel at real-world enterprise use cases. Command A is an agent-optimised and multilingual-capable model, with support for 23 languages of global business, and a novel hybrid architecture balancing efficiency with top of the range performance. It offers best-in-class Retrieval Augmented Generation (RAG) capabilities with grounding and tool use to automate sophisticated business processes. These abilities are achieved through a decentralised training approach, including self-refinement algorithms and model merging techniques. We also include results for Command R7B which shares capability and architectural similarities to Command A. Weights for both models have been released for research purposes. This technical report details our original training pipeline and presents an extensive evaluation of our models across a suite of enterprise-relevant tasks and public benchmarks, demonstrating excellent performance and efficiency.
核心贡献 · Key contributions
提出 Command A，一个 111B 参数的企事业大语言模型，采用新颖的混合架构，在效率与性能之间取得平衡。 Proposes Command A, a 111B parameter enterprise LLM with a novel hybrid architecture balancing efficiency and performance.
引入去中心化的后训练流程，通过并行专家模型与参数合并来综合多种能力。 Introduces a decentralized post-training pipeline with parallel expert models and parameter merging to combine multiple capabilities.
在 Taubench 等企业基准上取得最佳 RAG、工具使用和智能体性能，支持 23 种语言。 Achieves best-in-class RAG, tool use, and agentic performance on enterprise benchmarks like Taubench, supporting 23 languages.
展现出强大效率，仅需两块 A100/H100 GPU 即可部署，吞吐量可达每秒 156 个 token。 Demonstrates strong efficiency with a serving footprint of only two A100/H100 GPUs and up to 156 tokens/sec throughput.
在 MATH 等公开基准上表现优异，并在跨任务人类评估中超越更大规模的模型。 Shows excellent performance on public benchmarks such as MATH and outperforms larger models in human evaluations across tasks.
发布 Command A 和 Command R7B 的权重用于研究，并提供详细的训练流程和广泛评估。 Releases weights of Command A and Command R7B for research, with detailed training pipeline and extensive evaluations.
局限 · Limitations
模型合并评估是瓶颈，在生产环境中需要大量推理算力和时间。 Model merging evaluation is a bottleneck, requiring significant inference compute and time in production contexts.
专家碰撞可能降低性能；缓解依赖跨领域数据与一致性维护，限制了专家多样性。 Expert collisions can degrade performance; mitigation relies on cross-domain data and careful consistency, limiting expert diversity.
后训练通常会降低长上下文性能，仅能通过与具备长上下文能力的检查点合并来部分恢复。 Post-training generally degrades long-context performance, which is only partially recovered via merging with long-context capable checkpoints.
该方法针对企事业用例设计，在商业领域之外的一般性或创造性任务上效果可能有限。 The approach is tailored to enterprise use cases; effectiveness for general or creative tasks beyond business domains may be limited.
结果依赖内部奖励模型和合成数据；部分评估可能存在偏差且缺乏外部验证。 Results rely on internal reward models and synthetic data; potential biases and lack of external validation in some evaluations.