Falcon-180B released by TII on Hugging Face — hosting use restricted by license

AI Impact Summary

Falcon-180B is now publicly available via HuggingFace, marking the largest openly accessible LLM to date. It was trained on 3.5 trillion tokens across up to 4096 GPUs with SageMaker, and includes a chat-tuned variant, delivering strong open-model performance that rivals PaLM-2 on several benchmarks. The license allows commercial use but explicitly excludes hosting, which constrains how vendors can deploy it as a hosted service and may require separate licensing for SaaS offerings; teams can experiment and fine-tune offline but cannot provide hosted Falcon-180B endpoints under the same terms. The availability of quantization options (8-bit/4-bit) can reduce hardware needs for inference, influencing internal evaluation and open-model strategy decisions.