GGUF quantized versions of david-ar/20q, the world's smallest chat model.

This model was natively trained at 2-bit precision. All quantization levels above Q2_K are technically upscaled. Q2_K is the model's native precision.

All quantizations are lossless because the original weights are 2-bit integers (values 0-3). Q2_K is the only quantization level that doesn't waste bits.

general.architecture: twentyq
twentyq.block_count: 0
twentyq.embedding_length: 156
twentyq.attention.head_count: 156
twentyq.context_length: 20
twentyq.vocab_size: 1200

Zero transformer blocks. 156 attention heads. 20-token context window. The output projection layer (output.weight) contains the entire model.

These files require a runtime with twentyq architecture support, which does not currently exist in llama.cpp, ollama, or any other GGUF runtime. For inference, use the original model via the transformers library, or the live demo.