The text generation landscape has shifted dramatically in the last eighteen months. What was once a field dominated by closed APIs from OpenAI has become a genuinely competitive open-source ecosystem, and the numbers tell the story. Three of the largest openly licensed causal language models as of this writing are MPT-30B from MosaicML, XGen from Salesforce, and Falcon from TII UAE. Then Meta dropped Llama 2, which currently outperforms all other open-source large language models across benchmarks and comes with a license that explicitly permits commercial use.
That commercial angle is the quiet revolution here. Companies have been hesitant to route proprietary data through third-party APIs. Open-source models solve that problem directly. You keep your data in-house, fine-tune on your own domain, and avoid per-token inference costs that scale poorly as usage grows. The economic case was already compelling; Llama 2 simply removed the licensing ambiguity that kept legal teams nervous.
A useful distinction that often gets lost in the hype: not all instruction-tuned models are created equal. Causal language models like GPT-3 and Llama are trained to complete text, then adapted through reinforcement learning from human feedback to handle conversational prompts. Text-to-text models like Google’s FLAN-T5 series — currently state-of-the-art in that category — are trained on paired inputs and outputs from the start. Both live on the Hugging Face Hub, but they behave differently under the hood.
Hugging Face itself has skin in this game, having co-led the BigScience and BigCode initiatives that produced BLOOM and StarCoder. BLOOM, a causal language model trained on 46 languages, remains one of the more ambitious multilingual efforts in open source. The broader pattern is clear: the moat around closed models is shrinking, and the next battle won’t be about raw capability but about who can offer the cheapest, most customizable deployment path.
