Fetching from the wire…
01
02
03
04
05
06
07
08
Public story · 2026-08-22 · high
The DSL posts a 107 percent median speedup on MLX and 58 percent on native vLLM, a hedge against platform lock-in.
Why now: Axon's paper, arXiv 2608.19889, surfaced in the August 22 listing.
Axon compiles one model definition into standalone PyTorch, JAX, MLX and native vLLM code, per the paper posted to arXiv. It's a strongly typed DSL with Haskell-like syntax, built on the argument that open models shouldn't depend on one platform's implementation to run well.
Deployed as a native vLLM architecture with PagedAttention and KV-cache, already standard in production serving, Axon beat HuggingFace Transformers by a median 58%. That's the number most likely to show up in an actual deployment, not just a benchmark chart.
The full spread, across 467 benchmarking runs on models from 135M to 32B parameters, is uneven. Axon beat Transformers by a median 7% on plain PyTorch, 12% with Triton, 91% on JAX, and 107% on MLX.
That gap is the real story. A 7% gain on PyTorch says the reference code there was already close to tight. A 107% gain on MLX says its reference implementation was leaving over half its performance on the floor. Same model, same math, different headroom depending on which backend's reference code got real attention.
The paper doesn't say whether the speedups hold past 32B parameters, or what compiling costs in engineering time against writing PyTorch by hand.
Each link below shares sources, entities, or timing with this story.
Same source domain / Semantically similar
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.72).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.70).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.69).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.70).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.69).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.69).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.69).
Reported by the same outlet (arxiv.org); covers closely related ground (similarity 0.68).