Solo researcher releases Wave Field LLM, replacing dot-product attention with FFT wave convolution

A solo researcher has introduced Wave Field LLM, a new base completion model architecture that replaces standard O(N²) dot-product attention with FFT wave convolution. This approach allows for constant speed and memory usage during inference regardless of context length, enabling 128K context windows where standard attention typically runs out of memory.

• Training complexity is reduced to O(N log N), while inference achieves O(1) per token.
• The model family ranges from 130M to 1.5B parameters and was trained from scratch without RLHF or instruction fine-tuning.
• On a Mac laptop CPU, the model achieves over 80 tokens per second without requiring a GPU.
• Benchmarks on H100 show it is 21.8x faster and uses 5.3x less memory than standard attention at 32K context.
• In zero-shot evaluations on DCLM CORE, the 130M model outperformed GPT-2 124M, scoring 46.8% average compared to 26.5%, with significant gains on PIQA (61.7% vs 50.0%) and ARC Easy (43.8% vs 25.0%).

The author is seeking independent testing and feedback on benchmark methodology, quality comparisons with other ~130M base models, and potential issues such as harmful outputs.