--- title: "Cohere Inc. — 科希尔" source: "https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94" wiki: "systems-analysis.info/int" article: "Cohere_Inc._—_科希尔" language: "zh" categories: - "Category:Chinese" - "Category:Large language models" - "Category:LLM families" - "Category:Machine learning" revision_id: 1065 wiki_created_at: 2026-09-06T22:42:48Z wiki_modified_at: 2026-09-06T22:42:48Z downloaded_at: 2026-09-07T22:43:53Z --- # Cohere Inc. — 科希尔 **Cohere**是一家加拿大的AI技术公司，专注于为企业领域（Enterprise AI）开发大型语言模型（LLM）。该公司于2019年由Google Brain的前研究人员在多伦多创立，定位为OpenAI的主要替代品之一^{[\[1\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-wiki-cohere-1)}。 Cohere的主要重点是为企业提供安全、可定制和可扩展的AI解决方案。与许多面向消费市场的竞争对手不同，Cohere从一开始就专注于企业客户，这决定了其技术路线。 ## 历史与融资 Cohere由三位联合创始人于2019年创立：**艾丹·戈麦斯 (Aidan Gomez)**、**伊万·张 (Ivan Zhang)**和**尼克·弗罗斯特 (Nick Frosst)**。艾丹·戈麦斯是2017年里程碑式科研论文**《Attention Is All You Need》**的合著者之一，该论文向世界介绍了Transformer架构^{[\[1\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-wiki-cohere-1)}。2022年12月，YouTube前首席财务官马丁·科恩（Martin Kon）成为公司总裁兼首席运营官^{[\[1\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-wiki-cohere-1)}。公司在吸引投资方面取得了令人瞩目的成绩： - 2023年，在C轮融资中筹集了**2.7亿美元**，公司估值达到21亿至22亿美元^{[\[2\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-mostecosystem-funding-2)}。 - 2024年，公司额外筹集了**5亿美元**，估值增至55亿美元^{[\[3\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-uniteai-funding-3)}。 Cohere的主要投资者包括PSP Investments、Cisco Systems、AMD、Fujitsu、NVIDIA、Salesforce Ventures、Oracle和Google Cloud^{[\[1\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-wiki-cohere-1)[\[2\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-mostecosystem-funding-2)}。 ## 技术栈 ### Retrieval-Augmented Generation (RAG) - 检索增强生成 Cohere方法的核心技术是**检索增强生成 (Retrieval-Augmented Generation, RAG)**。RAG允许模型实时访问外部知识库，这将其回答“植根”于事实数据，减少“幻觉”并确保信息的时效性^{[\[4\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-4)}。**Command R**和**Command A**模型专门为RAG系统中的多步骤工具使用进行了优化^{[\[5\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-servernews-command-r-5)}。 ### 多语言能力 Cohere特别注重多语言能力。公司开发的模型使用母语者数据进行训练，确保了高质量的性能。例如，开源模型**Aya**支持包括俄语在内的100多种语言，而**Command**系列则支持超过10种全球主要语言。 ### 安全性与部署 Cohere提供灵活的部署选项，包括SaaS平台、私有云部署以及与第三方平台（Amazon SageMaker、Google Vertex AI、Azure）的集成。公司特别注重企业数据的安全性，提供选择退出数据用于训练的选项，并符合SOC 2 Type II标准^{[\[6\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-enterprise-data-commitments-6)}。 ## 产品与模型系列 Cohere开发了多种类型的模型，专为语言处理的不同阶段设计。 ### Command (Generative models) - Command (生成模型) **Command**是其旗舰生成模型系列。 - **Command A**: 最高效的模型（2025年），拥有**1110亿参数**。支持**256,000个词元**的上下文窗口，仅需两个GPU即可运行，而竞争模型则需要多达32个^{[\[7\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-habr-command-a-7)}。其生成速度高达每秒156个词元，比GPT-4o快1.75倍^{[\[8\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-dsmedia-command-a-8)}。 - **Command R+**: 强大的模型，拥有1040亿参数和128,000个词元的上下文窗口。性能超越许多同类模型，接近GPT-4 Turbo的水平^{[\[9\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-vc-future-command-r-9)}。 - **Command R**: 拥有350亿参数的模型，专为高精度对话任务设计^{[\[10\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-aigenom-command-r-10)}。 ### 专用模型 - **Aya**: 支持超过100种语言的开源多语言模型。由来自119个国家的3000多名研究人员合作开发^{[\[11\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-computerra-aya-11)}。 - **Embed**: 用于创建文本向量表示的模型。**Embed 4**支持多模态和128,000个词元的上下文窗口^{[\[12\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-habr-embed4-12)}。 - **Rerank**: 用于提升RAG系统中搜索质量的模型。 ## 合作伙伴与集成 Cohere积极与大型科技公司发展合作伙伴关系，包括： - **Oracle:** 将Cohere技术集成到Oracle Fusion Cloud和NetSuite中^{[\[13\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-erp-today-oracle-13)}。 - **Salesforce:** 将聊天功能嵌入Salesforce产品中^{[\[1\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-wiki-cohere-1)}。 - **Fujitsu:** 合作开发日语语言模型Takane^{[\[14\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-14)}。 - **McKinsey:** 在将生成式AI集成到企业运营方面进行合作^{[\[1\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-wiki-cohere-1)}。 ## 各行业应用 - **教育:** Cohere被用于创建课程计划、分类任务和支持多语言项目^{[\[15\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-arxiv-rag-education-15)[\[16\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-ieee-online-learning-16)}。 - **医疗保健与科学:** 模型被用于制定科研协议、分析医疗数据和支持研究^{[\[17\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-bmcmededuc-protocols-17)}。 - **企业解决方案:** 自动化文档工作流程、数据分析、撰写营销文案和客户服务^{[\[5\]](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_note-servernews-command-r-5)}。 ## 参见 - 大型语言模型 - ## 文献 - Cohere Team (2025). *Command A: An Enterprise‑Ready Large Language Model*. arXiv:2504.00698. - Üstün, A.; et al. (2024). *Aya Model: An Instruction Finetuned Open‑Access Multilingual Language Model*. arXiv:2402.07827. - Singh, S.; et al. (2024). *Aya Dataset: An Open‑Access Collection for Multilingual Instruction Tuning*. arXiv:2402.06619. - Cohere Team (2024). *Command R & Command R+: Technical Overview*. PDF. - Gao, Y.; et al. (2024). *Retrieval‑Augmented Generation for Large Language Models: A Survey*. arXiv:2312.10997. - Yu, H.; et al. (2024). *Evaluation of Retrieval‑Augmented Generation: A Survey*. arXiv:2405.07437. - Zhu, W.; et al. (2025). *PSC: Extending Context Window of Large Language Models via Phase Shift Calibration*. arXiv:2505.12423. - Gao, Y.; et al. (2024). *Fine‑Tuning vs. Retrieval‑Augmented Generation for Less Popular Entities*. arXiv:2403.01432. - Yu, L.; et al. (2025). *Synergizing RAG and Reasoning: A Systematic Review*. arXiv:2504.15909. - Wang, Y.; et al. (2025). *Universal Embeddings for Multimodal Multilingual Retrieval*. arXiv:2506.18902. ## 注释 1. ↑ ^{[1.0](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-wiki-cohere_1-0)} ^{[1.1](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-wiki-cohere_1-1)} ^{[1.2](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-wiki-cohere_1-2)} ^{[1.3](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-wiki-cohere_1-3)} ^{[1.4](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-wiki-cohere_1-4)} ^{[1.5](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-wiki-cohere_1-5)} “Cohere (company)”. 在*维基百科*中. [1] 2. ↑ ^{[2.0](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-mostecosystem-funding_2-0)} ^{[2.1](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-mostecosystem-funding_2-1)} “估值超21亿美元的AI初创公司Cohere融资2.7亿美元”. *Most Ecosystem*. [2] 3. [↑](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-uniteai-funding_3-0) “Cohere融资5亿美元”. *Unite.AI*. [3] 4. [↑](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-4) “Rethinking the RAG vs Fine-tuning Debate”. *arXiv*. [4] 5. ↑ ^{[5.0](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-servernews-command-r_5-0)} ^{[5.1](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-servernews-command-r_5-1)} “Cohere Command R：它是什么以及如何使用？”. *ServerNews*. [5] 6. [↑](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-enterprise-data-commitments_6-0) “Enterprise Data Commitments”. *Cohere Docs*. [6] 7. [↑](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-habr-command-a_7-0) “Cohere发布Command A——当今最高效的商业AI模型”. *Habr*. [7] 8. [↑](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-dsmedia-command-a_8-0) “Cohere发布低成本AI模型，仅需两块GPU”. *DSmedia*. [8] 9. [↑](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-vc-future-command-r_9-0) “面向企业的LLM：Cohere的Command R能做什么”. *VC.ru*. [9] 10. [↑](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-aigenom-command-r_10-0) “Cohere Command R”. *AIGenom*. [10] 11. [↑](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-computerra-aya_11-0) “Cohere for AI推出多语言AI ‘Aya’”. *Компьютерра*. [11] 12. [↑](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-habr-embed4_12-0) “Cohere推出Embed 4——其最强大的向量表示创建模型”. *Habr*. [12] 13. [↑](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-erp-today-oracle_13-0) “Oracle, Salesforce Ventures and Nvidia invest in Cohere”. *ERP Today*. [13] 14. [↑](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-14) “Fujitsu and Cohere to Partner on Developing Generative AI for the Japanese Language”. *Fujitsu Press Releases*. [14] 15. [↑](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-arxiv-rag-education_15-0) “Evaluating the Practicality of Retrieval-Augmented Generation in Student-Facing Educational Technology”. *arXiv*. [15] 16. [↑](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-ieee-online-learning_16-0) “A Confusion Classification System for Online Learning Based on Multimodal Information”. *IEEE Xplore*. [16] 17. [↑](https://systems-analysis.info/int/Cohere_Inc._%E2%80%94_%E7%A7%91%E5%B8%8C%E5%B0%94#cite_ref-bmcmededuc-protocols_17-0) “Using large language models to streamline the creation of research protocols”. *BMC Medical Education*. [17]