TEL AVIV — doubleAI announced Maxwell on October 19, 2026, and is accepting early-access requests for an AI system designed to make software execute more efficiently. Its Tel Aviv-datelined announcement describes a system that adjusts machine instructions while a program runs, with applications across CPUs and GPUs.
The Next Web's October 19 reporting also reported the launch and identified doubleAI as a Tel Aviv company. The lab also lists current engineering and research roles in neighbouring Ramat Gan. The development gives the local AI sector a product launch focused on the computing work beneath an application.
Optimizing the program after it starts
doubleAI calls Maxwell a code execution model. In the company's description, it observes the hardware, data and behaviour of running software, finds bottlenecks and substitutes a faster implementation underneath. The source files remain as their authors wrote them. The company presents Maxwell as complementary to coding assistants: an assistant can help produce a program, while Maxwell targets its execution.
The proposed value is easier to understand through an engineering question. Once a program is correct, how much unnecessary work does it still perform on a particular machine? Maxwell's commercial proposition is to find and remove that work repeatedly, as the conditions in which the program runs change.
The proof claim is central
The Maxwell technical explanation says replacements come with machine-checkable proofs that they preserve the original program's behaviour. It describes specializing code for observed conditions, then using a guard to return execution to the original version when an assumption no longer holds.
That conditional structure matters. A faster implementation may be valid for particular inputs, types or numerical ranges; the condition is part of the argument that permits the substitution. The company illustrates the idea with a deliberately simple pair of functions and an explicit bound on the input.
These are doubleAI's published claims about its verification system. A useful assessment will require the implementation, the assumptions captured by its formal definitions and evidence about how it behaves on a customer's actual software. The launch announcement alone does not establish independent validation of every supported workload.
A specific test: generating 4K video
A company research report dated October 19 offers a more concrete example. Working with LTX's LTX-2.5 video pipeline, doubleAI reports that generation of a five-second 4K clip took about 134 seconds after optimization, compared with 250 seconds before it, using two Nvidia B200 GPUs. That is approximately a 1.87-times speedup in the reported setup.
The report describes several kinds of changes: reducing communication between GPUs, reducing the overlap between image tiles to avoid repeated computation and improving work inside the video decoder. It links the optimized implementation as open source, giving engineers a route to inspect the changes.
Output quality is an important part of that example. doubleAI reports bit-identical audio, while the video comparison scores 41 dB on peak signal-to-noise ratio, a measure of pixel similarity. The company describes the videos as visually indistinguishable; it does not report bit-identical video frames.
The result is evidence supplied by the developers for a specified pipeline and hardware configuration. Its value for another workload depends on what is being computed, how performance is measured and which output differences are acceptable.
Research preceded this week's release
Maxwell's public research history extends beyond the new early-access announcement. In a March 5 report on GPU graph algorithms, doubleAI described generating specialized kernels and checking them with algorithm-specific verifiers. The report discusses why ordinary output comparisons can be inadequate for algorithms that permit several valid answers.
It also describes efforts to stop an optimization system from exploiting its evaluation process, including hidden test inputs and isolated evaluation environments. Those details help explain the engineering problem behind the headline: the system needs a reliable way to recognize a legitimate improvement before speed becomes a useful objective.
The March work is background to the October launch. Its specialized tests should be read in their own context, alongside the new runtime product's separate claims about proofs of equivalent behaviour.
From laboratory results to broader use
doubleAI's team page names Amnon Shashua as co-founder and CEO and Shai Shalev-Shwartz as co-founder and CTO. Investor Bessemer Venture Partners describes the company's initial focus as low-level code optimization, placing the launch within an existing research and investment strategy.
TLVNY's reading is that the next test is operational: how broadly Maxwell can be integrated, how much work it spends finding improvements and how engineers can inspect the conditions behind each change. Early access creates an opportunity to answer those questions on software beyond the published demonstrations.
The product site currently invites requests for access. For teams interested in trying it, the meaningful comparison will pair execution time and computing cost with correctness requirements on the same workload. Maxwell's launch puts that combination—performance and a reason to trust the change—at the centre of doubleAI's product pitch.
