Meta Releases Llama 3.3 With Improved Multilingual Capabilities and 128K Context

## The Open-Source AI Race Accelerates Meta released Llama 3.3 this week, making the weights freely available for research and commercial use under its custom open license. The model comes in 8B, 70B, and 405B parameter sizes, with the 70B variant delivering performance that benchmarks competitively against closed models several times its size on key tasks. ## Multilingual Improvements Llama 3.3's most significant upgrade is multilingual performance. The model was trained on a substantially expanded multilingual corpus covering 30 languages, with particular improvements in European and Southeast Asian languages. On the FLORES+ multilingual benchmark, the 70B variant scores 12 points higher than Llama 3.1 on average across all evaluated languages. ## Extended Context The 128K token context window (up from 32K in Llama 3.1) enables new use cases: full legal contract review, multi-chapter document analysis, and long-session coding assistance without context truncation. Meta notes that quality is maintained throughout the full context window, addressing a criticism of earlier long-context models that degraded toward the end of their context. ## Deployment and Fine-Tuning The open-weights release enables organizations to fine-tune Llama 3.3 on proprietary data without routing sensitive information through third-party APIs. This is particularly valuable for healthcare, legal, and financial services companies with strict data residency requirements. AWS, Azure, and Google Cloud all announced same-day support for managed Llama 3.3 deployment. ## Competitive Context Llama 3.3's 70B model is positioning itself as the go-to choice for organizations that want near-frontier performance with local deployment control. Mistral's Mistral Large 2 and Alibaba's Qwen 2.5 are the primary competitors in the open-weights segment.