doubleAI launches Maxwell, a model that rewrites programs' machine code while they run

doubleAI, the Tel Aviv lab whose co-founder and CEO is Amnon Shashua, on Monday announced what it calls a new category of model: Code Execution Models (CEMs). The first one, named Maxwell, opened for early access the same day. According to the company's press release, distributed through Business Wire, the idea is to leave source code alone and instead have a model rewrite the machine instructions underneath a program while it runs, tuned to the hardware, the data and the usage pattern it actually sees.

What doubleAI says Maxwell does

Most AI coding tools work on the developer's side of the stack: they read and write source code. doubleAI positions CEMs on the other side. In its description, a CEM takes three inputs (the running program, the hardware it runs on, and the data flowing through it) and produces machine instructions rather than code a human would review.

• Runtime, not repository. Maxwell rewrites the binary beneath the code continuously as conditions change. The source code is not modified.
• CPUs and GPUs. The release says Maxwell works on both, without naming specific chips or vendors.
• Proof of equivalence. doubleAI says Maxwell formally and mathematically proves that each optimization leaves the program's behavior unchanged.
• Compounding. Each proven, faster version becomes the starting point for the next round of optimization.

The numbers, and where they come from

The release claims that on real-world workloads Maxwell made data pipelines more than 38 times faster and AI inference kernels more than 4.5 times faster. These are doubleAI's own figures. The announcement does not describe the baselines, the hardware, the workloads or the methodology, and no independent evaluation is available yet. Speedups of that size usually depend heavily on how unoptimized the starting point was, so they say little about what a given team should expect.

The release does not mention pricing, name customers, or define what early access involves. In the release's company description, doubleAI says it has raised a $200 million Series A co-led by Lightspeed Ventures and Bessemer Venture Partners, with participation from Dell Technologies Capital, Greenfield, NVIDIA, Pitango and BRM.

Why the claim is hard, and what to check

Rewriting compiled code at runtime is not new in itself: JIT compilers and profile-guided optimization have done versions of it for decades. What doubleAI is adding is a learned model choosing the transformations, combined with a formal guarantee that behavior is preserved. That guarantee is the part that matters most and the part that is hardest to evaluate from a press release. Formal equivalence proofs are only as strong as their specification: what counts as "behavior" (floating-point results, memory ordering, timing, side effects) determines whether an optimization that passes the proof is actually safe to ship. The release does not say which of those properties the proofs cover.

Teams evaluating Maxwell would reasonably want to know how it handles numerical precision in inference kernels, how much overhead the runtime monitoring adds, how it interacts with existing compilers and kernel libraries, and what happens when the data or hardware shifts after an optimization has been applied.

Why it matters for people building on AI

For most Promptea readers, inference cost reaches them as a per-token API price set by a provider. Work like this sits below that layer. If runtime optimization of inference kernels delivers real, verifiable gains, it is one of the routes by which serving costs fall, and eventually prices with them. Teams that run their own models or heavy data pipelines are the more direct audience. Nothing about how you prompt or which model you pick changes today; this is an announcement to track until independent results appear.