doubleAI's Maxwell Claims 38x on Data Pipelines and a Formal Proof for Every Rewrite, in Controlled Access

doubleAI's product page for Maxwell, which the company calls "the first Code Execution Model," claimed as of Oct. 20 a 38x speedup on data processing pipelines and 4.25x on AI inference kernels. It is in "controlled availability," and the page says a request for early access guarantees neither acceptance nor a release date.

According to the Maxwell page, the model runs alongside a program, watches how it handles data on a particular machine, and rewrites the underlying binary to run faster, leaving the source code unchanged. It targets CPUs and GPUs. The page states that "every improvement is formally proven equivalent to the original code."

The featured results, all from doubleAI

The page lists four named results. None carries an independent measurement, and the page gives no price.

Workload	Hardware	Before	After
LTX-2.5, 4K video, one clip	Two B200 GPUs	250 s	134 s
Whisper, 180 transcriptions	RTX 4090	106.23 s	49.82 s
NetworkX, 15 runs	CPU	61.44 s	17.50 s

The page describes the LTX-2.5 result as 1.87x faster. It also says Maxwell's cuGraph drop-in replacement averages 3.6x faster and that the model beats NVIDIA's PyTorch baselines on 90 percent of the Blackwell kernels in SOL-ExecBench.

1. Data processing pipelines (headline)38 x
2. AI inference kernels (headline)4.25 x
3. cuGraph replacement (average)3.6 x
4. LTX-2.5 4K video1.87 x

Source: doubleAI, heymaxwell.ai, accessed 2026-10-20

The 38x figure is a headline for the category. The page does not tie it to one of the three named workloads, and the largest named gain on the table is the NetworkX run, 61.44 seconds to 17.50.

What the SOL-ExecBench claim measures

SOL-ExecBench, described in an arXiv paper submitted March 19, 2026, has 235 GPU kernel problems drawn from 124 AI models and targets NVIDIA Blackwell GPUs. Its authors, who include Christos Kozyrakis and Humphrey Shi, argue that benchmarks scoring speedup over software implementations give optimizers a moving target. They score kernels against analytically derived hardware limits instead.

Beating NVIDIA's PyTorch baselines on 90 percent of kernels is therefore the kind of relative result the benchmark's authors say is less informative than how close a kernel comes to the hardware bound. The Maxwell page, as fetched, does not give a SOL Score.

What doubleAI has claimed before

Runtime claims on small workloads are easy to repeat and hard to compare; the site's Qwen3-4B-Deslop piece covers a self-measured result on a laptop, with the damage rate counted.

On March 3, 2026, doubleAI said its WarpSpeed system had rewritten CUDA kernels in NVIDIA's cuGraph library for an average 3.6x speedup over versions NVIDIA's engineers had refined for a decade, according to Techtime. Techtime said every tested kernel improved and more than half gained over 2x, with the code posted on GitHub. Its author wrote that the gains would need to hold up under community scrutiny, and Techtime reported no independent verification.

The same 3.6x figure now appears on the Maxwell page as the cuGraph average. Techtime names Amnon Shashua as doubleAI's chief executive and co-founder and Shai Shalev-Shwartz as chief technology officer and co-founder.

The site has covered another proof-style claim: OpenAI's Quasi-Riemann paper rested on a document no journal had checked. Maxwell's equivalence proofs sit in the same position. The Terminal found no proof method, checker or sample proof on the page, and the claim covers every rewrite.

What would settle it is a published equivalence proof for one rewritten binary, which doubleAI has not released.