doubleAI launches Maxwell, a model that rewrites running code and proves the rewrite equivalent

doubleAI, a Tel Aviv startup co-founded by Mobileye CEO Amnon Shashua, launched Maxwell on 19 October 2026. The company calls it the first “code execution model”. The idea: instead of helping someone write code, the model sits next to a program while it runs and makes that program faster on the hardware and data it actually sees.

The Wall Street Journal reported the launch. We could not read the paywalled story ourselves, so the WSJ details below come from Quartz’s summary of it. Israel’s Calcalist and SiliconANGLE reported the launch the same day alongside the company’s funding.

What Maxwell is supposed to do

In doubleAI’s description, Maxwell “watches a program while it runs, rewrites the binary underneath to fit that exact machine and data, and formally proves every improvement equivalent to the original.” It works on both CPUs and GPUs, aimed at data pipelines and AI inference kernels.

The proof part is the unusual claim. doubleAI’s manifesto says its proofs “can be conditional”: when the condition in the proof does not hold at run time, “a very fast guard hands control back to the original code.” It also says the original and the rewrite “need not share a language”, so Python could be replaced by native code. The company goes further and claims a “complete formalisation of the Python programming language”, covering “ordinary, general-purpose Python as it exists in the real world.” That is a very large claim for a language as dynamic as Python, and doubleAI has not published it for outside review on the pages we read.

The numbers, all from doubleAI

The product page shows several side-by-side runs:

• LTX-2.5 4K video: 250 seconds per clip down to 134 seconds on two B200 GPUs, which doubleAI calls 1.87 times faster.
• Whisper-small speech-to-text: 180 transcriptions on one RTX 4090, 106.23 seconds down to 49.82 seconds, with “identical transcripts”.
• NetworkX betweenness centrality: 15 runs on a CPU, 61.44 seconds down to 17.50 seconds.
• GPU kernels: after one day of search, it says Maxwell beats NVIDIA’s PyTorch baselines on 90% of the Blackwell kernels in the SOL-ExecBench benchmark.

These are vendor demos. We have not run Maxwell, and there is no independent benchmark of it yet.

Per Quartz’s account of the WSJ story, Maxwell cut one early customer’s compute spending by 67%. LTX, a maker of open-weight multimodal models whose co-founder and CEO Zeev Farbman spoke to the WSJ, found that Maxwell made inference for its own model twice as fast.

How doubleAI wants to charge

Per the same account, doubleAI plans to bill customers a share of the savings. Shashua’s example: if a customer’s compute spending fell from $10 million to $6 million, doubleAI would take a share of the $4 million saved. He did not say what percentage.

Maxwell is in early access for now; Quartz reports a wider release is planned before the end of October.

The money is not new

doubleAI paired the launch with its funding disclosure: a $200 million Series A co-led by Lightspeed Venture Partners and Bessemer Venture Partners, with Dell Technologies Capital, NVIDIA, Greenfield, Pitango and BRM, per Calcalist. The company has about 40 employees. The round itself is not fresh: Globes reported it in November 2025, when the company was still in stealth as AA-I Technologies. SiliconANGLE, citing the WSJ, adds that doubleAI will use half the round to buy 1,000 NVIDIA B300 GPUs. Calcalist also notes that the company’s speedup figures “have not been independently verified”.

Why it matters

Compute is the biggest bill for many AI teams, and most of the effort to cut it goes into hand-tuned kernels and compilers. Maxwell’s pitch is to automate that work on live programs, with a proof attached so that “faster” does not quietly become “different”. If the proofs and guards hold up outside doubleAI’s own demos, that is a meaningful change in how optimization gets done. Until then, it is a strong claim from a well-funded company with a record of ambitious announcements, including its recent graph-theory result.

What to ask before you try it

• What exactly is proved, and under which conditions does the guard fall back to the original code?
• Does the savings share apply to measured savings on your workloads, and who measures them?
• What runs where: does Maxwell need access to your production data or binaries outside your environment?