doubleAI says its new Maxwell model can rewrite the code underneath a running program without changing its source files. Before a replacement runs, the company says, it must pass a machine-checkable proof that it preserves the original program’s behavior.
Maxwell observes execution, identifies bottlenecks and generates replacement binaries tailored to the hardware and data in front of it. That gives it a different role from an AI coding assistant proposing source-code edits for developers to review.
The potential audience includes AI builders trying to reduce inference and data-processing costs without rewriting their applications. doubleAI reports substantial CPU and GPU gains, including a roughly twofold improvement in a Whisper speech-recognition demonstration.
Maxwell is accepting requests for controlled early access, however, and is not generally available. Its performance figures come from vendor demonstrations. The company’s video-generation case study also raises a distinction between exact semantic equivalence and preserving output quality.
An AI Optimization Layer Between Code and Hardware
doubleAI calls Maxwell its first “Code Execution Model,” or CEM. The label describes a proposed deployment layer: AI sits between the program developers wrote and the instructions the machine executes.
According to the company, Maxwell watches a live program long enough to understand its bottlenecks and runtime conditions. It then generates new algorithms and machine instructions for that environment, verifies the replacement and swaps it in while execution continues. The source code stays unchanged, so a developer would not need to merge a performance patch to benefit.
This makes Maxwell complementary to coding agents. An LLM-based assistant might help write an application; Maxwell’s proposed job is to improve how that application executes on a particular CPU or GPU.
doubleAI explicitly compares the approach to just-in-time compilation, which generates executable code during runtime. The company claims Maxwell has greater freedom to replace the underlying algorithm, beyond translating or optimizing its existing implementation.
Execution context can change which implementation is efficient. A particular data shape, device or pattern of work may favor an approach that wouldn't be the best general-purpose choice. Maxwell’s proposition is to exploit that specificity without asking the application’s authors to maintain hardware-specific source changes.
“Code Execution Model” remains a product claim, not an established new industry category. The practical question is whether this combination of runtime observation, code generation and verification can work economically across unfamiliar workloads.
