Axon preprint reports framework-portable LLM specifications with faster inference

A preprint introduces Axon, a typed domain-specific language for defining LLM architectures and compiling them into standalone implementations for PyTorch, Triton-enabled PyTorch, JAX, MLX and vLLM. Across 467 inference benchmarks on models from 135 million to 32 billion parameters, the authors report median speedups over Transformers reference implementations, including 58% for native vLLM deployments using PagedAttention and KV-cache. The work remains a preprint and reports its own benchmark results.