Byte-level tokenization operates on raw bytes, enabling handling of any Unicode text without vocabulary gaps. Core idea: Instead of characters or subwords, tokenize at byte level (256 possible base tokens). Then apply BPE or other algorithms on bytes. Universal coverage: Any valid UTF-8 text can be tokenized, no unknown tokens ever. Handles emojis, rare scripts, code, everything. Used by: GPT-2, GPT-3, GPT-4 (byte-level BPE), CLIP text encoder. Implementation: Map bytes to printable characters for BPE processing, apply standard BPE on byte sequences. Trade-off: Non-ASCII characters use multiple bytes, so tokenization less efficient for non-English. CJK characters may use 3-4 bytes each. Comparison: Character-level has vocabulary per character (can be huge for Unicode), byte-level fixed at 256 base tokens. Benefits: No preprocessing needed, handles any input, robust to encoding issues. Multilingual consideration: Same model handles all languages but token efficiency varies significantly. Modern standard: Most production LLMs now use byte-level approaches for robustness.