Maxwell iconMaxwell

commercial Enterprise

doubleAI's code execution model: rewrites running CPU and GPU code to fit hardware and data, with each change formally proven equivalent.

agents cost reduction
38x Faster Data Pipelines (company figure)
$220M Raised (Nvidia, Dell, Lightspeed)
2023 Founded

Overview

Maxwell is what doubleAI calls "the first Code Execution Model." Instead of helping you write code, it works on code while it runs: it studies how a program uses the hardware and its data, rewrites the machine code to fit, and formally proves each change equivalent to the original. It targets CPU and GPU workloads such as data pipelines and AI inference kernels, without changes to your source code. doubleAI is led by Mobileye co-founder Amnon Shashua and launched Maxwell in October 2026 in controlled early access.

The Verdict

Who Should Use Maxwell?

Best For

  • AI labs and companies with large GPU inference bills
  • Teams running heavy Python data pipelines (pandas, graph analytics)
  • Organizations that want speed-ups without rewriting source code
  • Buyers who prefer paying from savings over a fixed license

Not Ideal For

  • Individual developers or small teams (no self-serve signup)
  • Anyone who needs it today: access is by application only
  • Teams that require independent benchmarks before buying
  • Light workloads where compute cost is not a real line item

What's Great

  • Works on the running program, so no source changes are needed
  • Every rewrite comes with a formal proof that behavior is unchanged
  • A guard falls back to the original code if the faster path's conditions stop holding
  • Covers both CPUs and GPUs, from pandas pipelines to Blackwell kernels
  • LTX cut 4K video generation from about 4 minutes to just over 2 per clip
heymaxwell.ai · LTX case study · The Next Web

Watch Out For

  • Early access only; acceptance is not guaranteed and there is no release date
  • All speed-up figures come from doubleAI, with no independent benchmarks yet
  • No public pricing; doubleAI plans to take an undisclosed share of savings
  • Rewrites production code at runtime without human review, which some teams won't allow
  • One named customer result (LTX) so far
The Next Web · SiliconANGLE

Pricing

Early Access
Share of savings
No public price; doubleAI plans to charge a cut of the compute you save. Apply on the site
View all features & details

How It Works

  • Analyzes a running program against its hardware and data
  • Rewrites machine code, not source code
  • Writes a formal equivalence proof for each change
  • Reverts to the original path when assumptions break
  • Runs without a human in the loop

Published Results (company-reported)

  • pandas: 38x on the PDS-H benchmark
  • Nvidia cuGraph: 3.6x faster on average as a drop-in
  • Blackwell: beat PyTorch baselines on most SOL-ExecBench kernels
  • Whisper-small on an RTX 4090: 106s down to 50s for 180 transcriptions
  • LTX-2.5 4K video on two B200s: 1.87x faster

Targets

  • Python first (doubleAI has formalized Python)
  • Examples in Rust and C
  • PyTorch code replaced with GPU kernels
  • CPUs (SIMD) and Nvidia GPUs

Company

  • doubleAI (AA-I Technologies Ltd.), founded 2023
  • CEO Amnon Shashua, co-founder of Mobileye
  • $200M Series A from Lightspeed and Bessemer, with Nvidia and Dell
  • Valued at over $1B

How It Compares

Feature Maxwell torch.compile Hand-tuned kernels
Works on Running machine code PyTorch source Source you rewrite
Source changes None Small Large
Correctness Formal proof per change Tests Tests and review
Scope CPU and GPU, any workload PyTorch models One kernel at a time
Availability Early access Free, open source Engineer time
Cost Share of savings Free Expert salaries

User Reviews

Loading reviews...