The Pulse
doubleAI Launches Maxwell to Rewrite Code While It Runs
doubleAI has opened early access to Maxwell, an AI system that rewrites machine instructions while programs run. The company claims speedups of more than 38 times for data pipelines and more than 4.5 times for AI inference kernels, while sa
AI.info Team · Written with AI · fact-checked ·
doubleAI says its new Maxwell system made data pipelines more than 38 times faster and AI inference kernels more than 4.5 times faster. The company announced the product on October 19, 2026, and opened it for early access. Maxwell’s pitch is distinct from a coding assistant: it changes the machine instructions beneath a program while that program is running, rather than editing the source code.
Maxwell changes the instructions beneath the program
The system observes a running program, the hardware executing it and the data moving through it. It then generates replacement binary instructions tailored to that environment, doubleAI says, while leaving the original source code untouched. The company says Maxwell works with both CPUs and GPUs and can continue optimizing as conditions change.
“AI has changed how software is written, but not how it runs,” said Prof. Amnon Shashua, doubleAI’s co-founder and CEO.
“Code Execution Models change that. CEMs, and Maxwell in particular, are agents on behalf of the hardware as opposed to code agents that are on behalf of the human user.”
Prof. Amnon Shashua, co-founder and CEO of doubleAI
Shashua describes the product as an agent working for the hardware, rather than for the programmer. In practical terms, the model sits between existing software and the processor, aiming to reduce the work needed to complete a task without requiring developers to rewrite their applications.
Formal proofs are the safety claim
Runtime changes raise a direct risk: a faster replacement could produce different results. doubleAI says Maxwell formally proves each optimization preserves the original program’s behavior, rather than relying only on tests. Its online technical explanation says the proof can depend on stated conditions; when those conditions do not hold, a guard returns execution to the original code.
The release does not provide benchmark procedures behind its broad speedup claims, such as the specific workloads, hardware configurations or comparison baselines. A separate doubleAI report on work with LTX gives a more detailed example: the company says Maxwell reduced a five-second 4K video-generation run from about 250 seconds to about 134 seconds on two Nvidia B200 GPUs. That result concerns one particular pipeline and is a company-reported measurement, not evidence that every workload will see a similar gain.
A new layer alongside coding agents
Maxwell is designed to complement tools that generate or edit source code. A coding agent writes a program for a user; Maxwell aims to improve how the resulting program runs on the machine it encounters. The distinction could matter for teams with large compute bills or performance-sensitive workloads, but the announcement does not describe customer access terms or how the system handles a range of production environments.
doubleAI says it raised a $200 million Series A co-led by Lightspeed and Bessemer Venture Partners. The release also names Dell Technologies Capital, Greenfield, NVIDIA, Pitango and BRM as investors. The company has positioned Maxwell as its first Code Execution Model, a category it says acts on behalf of hardware rather than a person using software.
Early access, without public pricing
The company’s own demonstrations include CPU and GPU workloads, and its Maxwell site invites developers to request early access. Those demonstrations and performance figures are presented by doubleAI; the launch materials do not set out independent evaluations or a full testing methodology. The release opens early access, but gives no price or general-availability date.