Claude's tokenizer has only 16K vocabulary, sparking speculation about softmax bottleneck

rickasaurus · x · 2026-08-14

Developers have discovered that the latest Claude models have a tokenizer vocabulary of only about 16K, far lower than Qwen3.5's 250K. This anomaly has sparked discussion, with theories suggesting Anthropic may be trying to remove the softmax bottleneck by reducing vocabulary size. A reproduction project confirmed that Claude 4.7+ has an even lower estimate of just 15K entries.

Original post →

More from Models

Models channel →