doubleAI introduces Maxwell, a model that runs inside code to speed it up

A model that runs inside your program

doubleAI has released an early-access version of Maxwell, a type of AI model that does something quite different from what most AI tools do today. Instead of writing code for humans, Maxwell runs inside programs themselves and rewrites them in real time to make them faster ¹.

For decades, the division of labor has been simple: humans write instructions, and computers carry them out. Coding assistants like GPT-6 Astra and tools such as OpenAI Codex have changed how code is written, but they still hand the finished product to the same ordinary computer to run ¹. Maxwell is built to step into that gap — between the code and the silicon.

How it works

Maxwell, which the company also calls a "code execution model" or CEM, watches data flow through a running program and spots slow spots, called bottlenecks ¹. When it finds one, it builds a new version of the code specifically for the machine it is running on . The original source code — the human-readable instructions — stays exactly as the programmer wrote it . Only the binary, the low-level code the processor actually runs, changes .

Because these changes happen without a human checking them, getting them wrong could break a program. To prevent that, Maxwell uses formal verification, a method that turns code into precise mathematical statements and then proves that a new version does exactly the same thing as the old one ¹. This is the same approach that recently helped produce complete formal proofs of major math problems .

The result, according to doubleAI, is much faster binaries that are proven to be semantically equivalent to the originals ¹. Two programs are semantically equivalent when, given any input, they always produce the same result and have the same effects on the world around them .

What this means for coding agents

doubleAI says Maxwell is not meant to replace coding agents ¹. They do different jobs: a coding agent writes the program, and Maxwell makes sure the machine runs it as efficiently as possible . Together, they would cover the two parts of the process that AI has touched so far — drafting and then executing code.

The company has made no public benchmark numbers or performance claims beyond its own description ¹. No outside testing has been reported.

Why it matters

Most AI models today stop at writing or editing code. Maxwell tries to go one step further by treating code as something that can evolve while it runs ¹. If the approach scales, it could change how software is deployed, letting programs adapt their own inner workings on the exact hardware they land on . Whether Maxwell can keep its promise of perfect correctness at scale remains to be seen.

This story reports on a trend for information only. It is not advice or encouragement to do any of it.

Key facts

• Maxwell is a 'code execution model' that runs inside programs and rewrites binary code in real time.
• It uses formal verification to mathematically prove that optimized code produces the same results as the original.
• The original source code stays unchanged; only the low-level binary is modified.
• doubleAI says Maxwell complements coding agents rather than replacing them.

Sources

1. Maxwell: The first Code Execution Model, by doubleAI — heymaxwell.ai
2. OpenAI Codex — en.wikipedia.org