Maxwell: A new kind of model that sits between code and hardware

Maxwell is a new class of "code execution models" that sits between source code and silicon, running inside programs to observe data flow and hardware characteristics and then emit machine- and data-specific optimized binaries while the program keeps running. It replaces hotspots with formally proven, semantically equivalent implementations without changing the original source. Demonstrations and benchmarks show large speedups: 38× faster data pipelines, 4.25× faster AI inference kernels, LTX-2.5’s 4K generation nearly twice as fast (250s → 134s), Whisper transcription on an RTX 4090 cut from 106.23s to 49.82s, and NetworkX betweenness centrality reduced from 61.44s to 17.50s. Research highlights include outperforming NVIDIA PyTorch baselines on 90% of Blackwell kernels, a 3.6× faster cuGraph replacement, and fully reworked pandas kernels for modern hardware.

The engine behind Maxwell is formal verification: every transformation is accompanied by a machine-checkable proof of semantic equivalence that covers effects, mutable state, exceptions and I/O. Python’s dynamic semantics have been fully formalized to permit these proofs; guarded runtime checks hand control back to the original code when preconditions don’t hold. The design leverages the Curry-Howard correspondence and proof-checking infrastructure so optimizations are deterministic and provably safe. Maxwell is framed as complementary to code-writing agents: they author programs, and Maxwell ensures they run optimally on the exact machine and data in front of it.