--- title: "Falcon (language model family) — 猎鹰模型" source: "https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B" wiki: "systems-analysis.info/int" article: "Falcon_(language_model_family)_—_猎鹰模型" language: "zh" categories: - "Category:Chinese" - "Category:Large language models" - "Category:LLM families" - "Category:Machine learning" revision_id: 2114 wiki_created_at: 2026-09-06T22:58:49Z wiki_modified_at: 2026-09-06T22:58:49Z downloaded_at: 2026-09-07T22:49:24Z --- # Falcon (language model family) — 猎鹰模型 **Falcon** 是一个开源大型语言模型 (LLM) 家族，由阿联酋阿布扎比的**技术创新研究所** (Technology Innovation Institute, TII) 开发^{[\[1\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-interfax-release-1)}。Falcon模型已成为推动普惠人工智能发展的重要贡献，并经常在 Hugging Face 的 Open LLM Leaderboard 等性能排行榜上名列前茅^{[\[2\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-hf-blog-falcon-2)}。该模型家族包括各种规模和专业化的模型，从可在消费级硬件上运行的紧凑版本，到能与顶尖科技公司产品相媲美的最大规模模型。Falcon 的主要特点包括其先进的架构、在高质量数据集 **RefinedWeb** 上的训练，以及主要采用的开放式 Apache 2.0 许可证^{[\[3\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-hf-transformers-doc-3)}。 ## 历史与发展 Falcon 模型的首个版本于2023年6月发布。2023年9月，**Falcon-180B** 模型问世，当时它成为全球最大、性能最强的开源 LLM，其参数数量超过了 Meta 的 Llama 2 70B^{[\[4\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-decoder-180b-review-4)[\[5\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-neurohive-180b-5)}。该家族的后续发展包括新一代模型和专业版本的发布： - **Falcon 2** (2024年)：第二代产品，功能有所提升，包括多模态版本 **Falcon 2 11B VLM** (Vision Language Model)^{[\[6\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-aetoswire-falcon2-6)}。 - **Falcon 3** (2024年12月)：最新一代模型，在14万亿个 token 上进行训练，具备增强的多模态功能，并针对包括笔记本电脑在内的轻量级硬件进行了优化^{[\[7\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-aetoswire-falcon3-7)[\[8\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-mediaoffice-falcon3-8)}。 - **专业模型**：发布了针对特定任务的定制模型，如 **Falcon Arabic** 和 **Falcon Mamba**。 | 模型 | 参数（十亿） | 主要特点 | 许可证 | |-------------------|--------------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|-------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------| | **Falcon-180B** | 180 | 第一代最大模型；在3.5万亿个 token 上训练；性能超越 GPT-3.5^{[\[4\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-decoder-180b-review-4)}。 | TII Falcon License 1.0 (有商业使用限制)^{[\[5\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-neurohive-180b-5)} | | **Falcon-40B** | 40 | 基础高性能模型；在1万亿个 token 上训练。 | Apache 2.0 | | **Falcon-7B** | 7 | 紧凑型模型，需要约15 GB GPU内存；适用于消费级硬件^{[\[2\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-hf-blog-falcon-2)}。 | Apache 2.0 | | **Falcon-1.3B** | 1.3 | 适用于资源受限设备的最小模型。 | Apache 2.0 | | **Falcon 2 11B** | 11 | 第二代；与 Llama 3 8B 和 Gemma 7B 竞争；存在多模态版本 (VLM)^{[\[6\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-aetoswire-falcon2-6)}。 | Apache 2.0 | | **Falcon 3** | N/A | 在14万亿个 token 上训练；多模态（文本、图像、音频、视频）；可在笔记本电脑上运行^{[\[7\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-aetoswire-falcon3-7)}。 | Apache 2.0 | | **Falcon Arabic** | 7 | 针对阿拉伯语（标准语和方言）的专业模型；采用 Falcon 3 架构^{[\[9\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-tii-falcon-arabic-page-9)}。 | Apache 2.0 | | **Falcon Mamba** | N/A | 基于 Mamba (SSM) 架构的实验性模型，取代了 Transformer 架构^{[\[10\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-pikabu-mamba-10)}。 | Apache 2.0 | Falcon家族主要模型 ## 架构与技术特点 ### Transformer 架构大多数 Falcon 模型基于“仅解码器”(decoder-only) 的 Transformer 架构构建。关键的架构决策包括： - **Multi-Query Attention (MQA)**：与标准的 Multi-Head Attention 中每个“头”都有自己独立的键值对 (key/value) 不同，在 MQA 中，所有注意力头共享同一组键和值。这显著减少了内存消耗并加快了推理速度，而不会造成明显的质量损失^{[\[2\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-hf-blog-falcon-2)}。 - **Rotary Positional Embeddings (RoPE)**：与其他现代 LLM 一样，使用 RoPE 对 token 的位置信息进行编码。 - **FlashAttention**：用于优化注意力机制的计算。 ### Mamba 架构 (State Space Model) **Falcon Mamba** 模型是一项创新，因为它摒弃了传统的 Transformer 架构，转而采用**状态空间模型 (State Space Model, SSM)**。Mamba 架构以线性方式处理数据序列，使其在处理极长上下文时效率显著提高，并且与 Transformer 相比需要更少的计算资源^{[\[10\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-pikabu-mamba-10)}。 ### 训练数据 Falcon 模型的训练基础是 TII 创建的高质量数据集 **RefinedWeb**^{[\[5\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-neurohive-180b-5)}。该数据集包含数万亿个从 Common Crawl 提取的 token，并经过严格的过滤和去重以提高数据质量。 - **Falcon-180B** 使用了一个包含 **3.5万亿个 token** 的扩展数据集，其中约85%来自 RefinedWeb，其余部分包括精选的书籍、对话和代码数据^{[\[4\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-decoder-180b-review-4)}。 - **Falcon Arabic** 在一个高质量的、原生的（非翻译的）阿拉伯语数据集上进行训练，该数据集涵盖了现代标准阿拉伯语和多种地区方言^{[\[11\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-tii-news-falcon-arabic-11)}。 ## 专业模型 ### Falcon Arabic **Falcon Arabic** 是一款拥有70亿参数的模型，专门针对阿拉伯语进行了优化。它在阿拉伯语基准测试（Open Arabic LLM Leaderboard）中表现出色，能够理解**现代标准阿拉伯语 (MSA)** 和多种地区方言。这使得该模型能够为阿拉伯语用户提供具有文化意识的准确回答^{[\[9\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-tii-falcon-arabic-page-9)}。其性能甚至超过了规模大其10倍的模型^{[\[12\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-falconfoundation-arabic-perf-12)}。 ### 多模态能力 - **Falcon 2 11B VLM** 是该家族中首个能够同时处理文本和图像的多模态模型^{[\[6\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-aetoswire-falcon2-6)}。 - **Falcon 3** 显著扩展了这些能力，增加了对**视频和音频**的支持。预计在2025年1月将提供完整的语音模式^{[\[7\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-aetoswire-falcon3-7)}。 ## 性能与问题 ### 与竞争对手的比较 Falcon 模型一直表现出高性能。 - **Falcon-180B** 在大多数学术基准测试（如 MMLU、HellaSwag 和 LAMBADA）上超过了 GPT-3.5 和 Llama 2 70B，但仍次于 GPT-4^{[\[4\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-decoder-180b-review-4)}。 - **Falcon 2 11B** 的性能与 Meta Llama 3 8B 和 Google Gemma 7B 持平或更高^{[\[6\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-aetoswire-falcon2-6)}。 - **Falcon 3** 在发布时，在其同等规模的模型中，位居 Hugging Face 全球排行榜第一名^{[\[7\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-aetoswire-falcon3-7)}。 ### 局限性与问题 - **多语言质量**：大部分训练数据是英语^{[\[13\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-13)}。因此，模型在包括俄语在内的其他语言上的表现质量可能显著较低。^{[\[14\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-aetoswire-falcon-h1-rus-14)}。 - **幻觉**：与所有 LLM 一样，Falcon 模型也容易产生不准确或虚构的信息（幻觉），这要求在关键应用中使用时需谨慎对待^{[\[15\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-habr-hallucinations-15)}。 - **许可证限制**：尽管大多数模型在 Apache 2.0 许可下发布，但旗舰模型 Falcon-180B 拥有自己的 **TII Falcon LLM License**，该许可规定，当商业使用收入超过100万美元时，需要支付版税，这限制了其在商业领域的应用^{[\[5\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-neurohive-180b-5)[\[16\]](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_note-hf-180b-discussions-16)}。 ## 外部链接 - Falcon LLM 官方网站 - TII 在 Hugging Face 上的主页 - GitHub 仓库 ## 参考文献 - Ainslie, J. et al. (2023). *GQA: Training Generalized Multi‑Query Transformer Models from Multi‑Head Checkpoints*. arXiv:2305.13245. - Almazrouei, E. et al. (2023). *The Falcon Series of Open Language Models*. arXiv:2311.16867. - Dao, T. et al. (2022). *FlashAttention: Fast and Memory‑Efficient Exact Attention with IO‑Awareness*. arXiv:2205.14135. - Ding, Y. et al. (2024). *LongRoPE: Extending LLM Context Window Beyond 2 Million Tokens*. arXiv:2402.13753. - Fedus, W.; Zoph, B.; Shazeer, N. (2021). *Switch Transformers: Scaling to Trillion Parameter Models with Simple and Efficient Sparsity*. arXiv:2101.03961. - Gu, A.; Dao, T. (2023). *Mamba: Linear‑Time Sequence Modeling with Selective State Spaces*. arXiv:2312.00752. - Penedo, G. et al. (2023). *The RefinedWeb Dataset for Falcon LLM: Outperforming Curated Corpora with Web Data, and Web Data Only*. arXiv:2306.01116. - Peng, B. et al. (2023). *YaRN: Efficient Context Window Extension of Large Language Models*. arXiv:2309.00071. - Shazeer, N. (2019). *Fast Transformer Decoding: One Write‑Head is All You Need*. arXiv:1911.02150. - Su, J. et al. (2021). *RoFormer: Enhanced Transformer with Rotary Position Embedding*. arXiv:2104.09864. ## 注释 1. [↑](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-interfax-release_1-0) “阿联酋推出大型语言模型Falcon 2”。*国际文传电讯社*。[1] 2. ↑ ^{[2.0](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-hf-blog-falcon_2-0)} ^{[2.1](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-hf-blog-falcon_2-1)} ^{[2.2](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-hf-blog-falcon_2-2)} “Falcon: The \\T-shirt-sized\\ 7B and 40B models that are democratizing the LLM landscape”. *Hugging Face Blog*. [2] 3. [↑](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-hf-transformers-doc_3-0) “Falcon Model”. *Hugging Face Transformers documentation*. [3] 4. ↑ ^{[4.0](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-decoder-180b-review_4-0)} ^{[4.1](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-decoder-180b-review_4-1)} ^{[4.2](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-decoder-180b-review_4-2)} ^{[4.3](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-decoder-180b-review_4-3)} “Falcon 180B open-source language model outperforms GPT-3.5 and Llama 2”. *The Decoder*. [4] 5. ↑ ^{[5.0](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-neurohive-180b_5-0)} ^{[5.1](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-neurohive-180b_5-1)} ^{[5.2](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-neurohive-180b_5-2)} ^{[5.3](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-neurohive-180b_5-3)} “Falcon 180B：全球最大的开源语言模型”。*Neurohive*。[5] 6. ↑ ^{[6.0](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-aetoswire-falcon2_6-0)} ^{[6.1](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-aetoswire-falcon2_6-1)} ^{[6.2](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-aetoswire-falcon2_6-2)} ^{[6.3](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-aetoswire-falcon2_6-3)} “Falcon 2：阿联酋技术创新研究所发布新系列AI模型，性能超越Meta的Llama 3”。*AETOSWire*。[6] 7. ↑ ^{[7.0](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-aetoswire-falcon3_7-0)} ^{[7.1](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-aetoswire-falcon3_7-1)} ^{[7.2](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-aetoswire-falcon3_7-2)} ^{[7.3](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-aetoswire-falcon3_7-3)} “Falcon 3：阿联酋技术创新研究所推出全球最强大的小型AI模型”。*AETOSWire*。[7] 8. [↑](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-mediaoffice-falcon3_8-0) “Technology Innovation Institute launches Falcon 3 model to enhance access to AI through light infrastructures”. *Abu Dhabi Media Office*. [8] 9. ↑ ^{[9.0](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-tii-falcon-arabic-page_9-0)} ^{[9.1](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-tii-falcon-arabic-page_9-1)} “Falcon Arabic”. *FalconLLM TII*. [9] 10. ↑ ^{[10.0](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-pikabu-mamba_10-0)} ^{[10.1](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-pikabu-mamba_10-1)} “Falcon Mamba——语言模型发展中无需注意力机制的新一步”。*Pikabu*。[10] 11. [↑](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-tii-news-falcon-arabic_11-0) “Middle East's Leading AI Powerhouse TII Launches Two New AI Models”. *TII News*. [11] 12. [↑](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-falconfoundation-arabic-perf_12-0) “Middle East's leading AI powerhouse, TII,launches two new AI models”. *Falcon Foundation*. [12] 13. [↑](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-13) Almazrouei, Ebtesam, Hamza Alobeidli, Abdulaziz Alshamsi, Alessandro Cappelli, Ruxandra Cojocaru, Mérouane Debbah, Étienne Goffinet, et al. “The Falcon Series of Open Language Models.” arXiv, November 29, 2023. https://doi.org/10.48550/arXiv.2311.16867.[13] 14. [↑](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-aetoswire-falcon-h1-rus_14-0) “中东领先的人工智能巨头TII推出两款新AI模型”。*AETOSWire*。[14] 15. [↑](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-habr-hallucinations_15-0) “Falcon-180B：评测、启动与初体验”。*Habr*。[15] 16. [↑](https://systems-analysis.info/int/Falcon_(language_model_family)_%E2%80%94_%E7%8C%8E%E9%B9%B0%E6%A8%A1%E5%9E%8B#cite_ref-hf-180b-discussions_16-0) “Falcon 180B License Discussion”. *Hugging Face*. [16]