Model Architecture & Training MethodsMeta (Pagnoni et al.)Meta

Byte Latent Transformer

Byte Latent Transformer

Why it matters

Processes bytes directly, without a fixed vocabulary, using entropy-based dynamic patches; the first large-scale scaling of the tokenizer-free approach

Source — Primary sources first; secondary reports are labelled “Press”

Paperarxiv.orghttps://arxiv.org/abs/2412.09871

Added in Week 35, 2026 · See that week

Same week

Ilya Sutskever gives the NeurIPS 2024 Test of Time talkPressGoogle introduces Gemini 2.0OfficialLG AI Research open-sources EXAONE 3.5Official

Nearby on the same axis

Keller Jordan releases the Muon optimizerOfficialDeepSeek-V3 Technical ReportPaperNie et al. present LLaDAPaperNSA — Native Sparse AttentionPaper