Mathlabs

Foundation models that run real systems.

Language models learned from text. Ours learn from the live measurements of any system someone has to keep running: a network, a portfolio, a grid. They see a problem coming, explain it in plain terms, and tell the team how to fix it.

Inside the model, every step has a name. Raw signals become concepts: named problems that an engineer or an analyst already recognizes.

  1. See it coming

  2. Explain it

  3. Fix it

Returns

Four uses of our models, each with its own return.

MeasuredExpected rangeEach bar runs from 0 to 75%

Collections

Works out why a bill is unpaid, then takes the best next action for each account.

Measured over ten weeks against a randomized control group, at a mobile operator with more than two million subscribers.

  • Churn after an overdue bill is chased25–35% lower
  • Cost to collect15–30% lower
  • Bad debt written off10–15% lower
  • EBITDA+30–65M a year for every 10 billion in revenue

Network experience

Finds the network faults customers are about to feel, and says what to fix.

Expected range, from comparable operator deployments. Running at a tier 1 mobile operator.

  • Time to find the root cause and restore service40–60% lower
  • Care contacts caused by the network25–40% fewer
  • Churn in the affected cells10–15% lower

SIM-swap fraud

Catches an account takeover while it is still forming.

Expected range, from comparable fraud deployments. Running at a tier 1 mobile operator in Eastern Europe.

  • False alarms on genuine customers50–70% fewer
  • Fraud losses40–60% lower
  • Time to detect a takeoverMinutes where rules take hours

Production failures

Traces failures on the production line back to their cause.

Measured across root-cause studies led by our chief scientist for European manufacturers, from semiconductor wafers to paper.

  • Failure events traced to their root cause40–70%

Why a model for tables, and why now

Language models changed work with text. The same change is arriving for tables, and it is arriving fast.

Tables run the business
Credit, fraud, pricing, collections, network faults. The decisions with money and regulators attached are made from rows and columns, and language models were not built to read them.
2026 was the year it turned
SAP bought Prior Labs. Google, NVIDIA and Amazon each released a tabular foundation model. Fundamental raised a quarter of a billion dollars. The category is now real.
The race is on accuracy
Far less effort goes into the model for the buyer who has to explain a decision to a customer, an auditor or a regulator. That buyer is most of the market.

At the centre of it all is one model. We built it.

Able V1

Our foundation model for tables. Open source, and built to be read.

A tabular foundation model
It reads a table it has never seen and predicts in one pass, with no training run.
Open source
We are releasing the weights and the code for anyone to inspect and build on. It stays open.
Readable by design
Every prediction comes with the named concepts behind it.

One table, one pass, one reason

Able V1 reads a table it has never seen, predicts the missing value, and says why.

A table it has never seen. No training run.

Cell sectorIndoor signalLink failuresTraffic against forecastHandover failuresComplaints next week
A12−92 dBm698%1.1%No
B07−104 dBm2197%1.4%Yes
C03−95 dBm8101%6.8%Yes
D19−93 dBm434%0.9%Yes
E22−91 dBm799%1.2%No
F05−96 dBm596%1.0%No

Its prediction for the new row, with reasons

    An illustration with invented values. The concept names come from our work with mobile operators.

    Inside Able V1

    What it is, where it came from, and why we built it this way.

    Type
    Tabular foundation model. Reads a new table and predicts in one pass, with no training run.
    Lineage
    Started from TabICLv2, then retrained with our own priors and enough new components to be our own baseline.
    Trained on
    Synthetic tables generated from cause-and-effect maps, shaped by expert knowledge of each field.
    Built first for
    Telecom and financial services.
    Interpretability
    Built in. Every prediction comes with the concepts behind it.
    Licence
    Open source, under a permissive licence. It stays that way.
    Release
    Weights, code and a technical report, in preparation.

    Tabular foundation models are where language models were a few years ago. The idea has been proven, the big companies have arrived, and nobody yet owns the category.

    Able V1 is built to perform in line with the best open models, and to do better in the industries we know. We will publish the results. That is necessary, and it is not why we exist.

    We exist because a model that decides things should be readable.

    Most models in this field are built to win a benchmark. Ours is built to compete on the benchmark and then show its working. Interpretability is part of the model and part of every release, and it always will be.

    The model is open because a model you are asked to trust should be one you can inspect.

    Built to be read

    Three things the model can do that a score cannot.

    The model and its readable layer

    Real data Synthetic data Able V1 Readable layer sleeping cell coverage hole unwanted behaviour stopped A recommendation, with its reasons The team decides
    The readable layer sits beside the model. It names what the model is doing, and it can stop a behaviour before the decision is made.
    It names what it uses
    A readable layer separates what each neuron mixes together into single ideas, with names a specialist recognizes: a sleeping cell, a billing dispute, a tariff scare.
    h ≈ Σi ai di
    It shows its path
    Every prediction can be traced from the cells it read, through the concepts it formed, to the answer it gave.
    x → a → y
    It can be stopped
    The same layer watches the model while it works. Behaviour nobody wants is turned down before the decision is made.
    ai ← 0

    A growing vocabulary for real systems

    Language models started with words. Real systems have none, so each one we work on adds its own: named concepts a model can use and a person can check.

    From a mobile network. Select a concept to see its signature.

    Fig. 1. Each shape is one concept’s signature across twelve data sources: the further a point sits from the centre, the more that source defines the concept. These are the expert starting points, which an operator’s own data then refines.

    From a credit book: risk stories, January to March.

    Fig. 2. Each line is one risk story found in a book of European corporates. It rises when new evidence arrives and fades when none does. Stories are ordered by when they first appeared.

    The field

    Eight models that matter today, and how open each one is.

    ModelMakerWeightsFree for commercial use
    Able V1MathlabsOpen sourceYes
    TabICLv2InriaBSD-3-ClauseYes
    Mitra-v2AmazonApache-2.0Yes
    Kumo TabularNVIDIAOpenMDW-1.1Yes
    TabPFN-3.5Prior Labs, now SAPNon-commercialNo
    TabFMGoogleNon-commercialNo
    LimiX-2Stable AINon-commercialNo
    NEXUSFundamentalProprietaryNo

    Table 1. Licences as published by each maker in November 2026. The public release of Able V1 is in preparation.

    The model is free. This is what we sell.

    Open models win by being everywhere. The business is what companies need on top of them.

    Each company ownsits model, outright

    An operator’s modelIts data, its concepts, its actions
    A bank’s modelIts data, its concepts, its actions
    A manufacturer’s modelIts data, its concepts, its actions

    We sellwhat a company needs to run it

    Observability and guardrailsWatch the model while it works, and stop what it should not do
    Industry versionsAble V1 tuned for telecom, banking and industry, with their concept vocabularies

    Open to everyonefree, and it stays free

    Able V1The foundation model, with its readable layer
    Free model, wide adoption
    Anyone will be able to download Able V1, check it and build on it. Adoption is how a model becomes a standard, and it is how the best researchers find us.
    Paid for production
    A company that puts the model in charge of real decisions needs the industry version, the guardrails and someone accountable. That is the product.
    The company owns its model
    It runs inside the company’s own walls and is exclusive to it. Its data never leaves, and never enters ours.

    Selected work

    Deepest in telecom, where three assets cover experience, fraud and collections. The same method is at work in banking and industry.

    SectorPartnerWork
    Telecom A tier 1 mobile operator Network experience pre-emption. Ten named concepts, from sleeping cells to coverage holes, scored for every cell sector every fifteen minutes. Alerts are raised before complaints arrive.
    Telecom A tier 1 mobile operator in Eastern Europe SIM-swap fraud, caught while it forms. Our tabular foundation model and its interpretability layer turn SIM history into named fraud concepts. Every open case is re-scored as events arrive, or fail to arrive.
    Telecom A mobile operator with more than two million subscribers Collections, next best action. Diagnoses why a bill is unpaid, then launches the best action for each account, automatically or through an officer. Bad debt fell 10 to 15 percent, measured against a randomized control group.
    Banking A tier 1 European bank Credit spread concept monitor. Risk stories are discovered as they form and tracked as they spread from name to name. Exposed names that have not yet moved go on a watchlist.
    Manufacturing European producers, from semiconductor wafers to paper Root cause of production failures. Causal analysis of anomalies on the line, which uncovered the root cause of 40 to 70 percent of failure events.
    Engineering A global engineering company based in the UK Accident root-cause engine. Taken from mathematical prototype to production, including the discovery of concepts that people can understand.
    Banking A tier 1 bank Merchant fraud check. A concept-based analysis for onboarding merchants, built for cases where the information is partial and inconsistent.
    Telecom A telecom operator An operator-owned language model. A small model built for the operator’s domain, running on its own infrastructure.

    Table 2. Selected work by Mathlabs and its team.

    Team

    Six of our scientists and engineers hold a PhD in mathematics, physics or computation, and a seventh is completing one.

    Erez Raanan

    Erez Raanan

    CEO and co-founder

    Former global leader of McKinsey Digital and Analytics Ventures. MBA from Harvard Business School, BSc in electrical and computer engineering.

    LinkedIn
    Prodipto Ghosh

    Prodipto Ghosh

    Chief Scientist and co-founder

    Former lead scientist at McKinsey, with more than fifteen years in computational mathematics. B.Tech from IIT Kanpur, MBA in operations research from IIM Bangalore, Masters in Quantitative Finance from London Business School.

    LinkedIn
    Tomer Raanan

    Tomer Raanan

    Chief Technology Officer

    More than twenty years building data, AI and cloud platforms. Former engineering leader at monday.com, and VP Engineering and first employee at Otonomo. MBA and BSc in computer science from Tel Aviv University.

    LinkedIn

    Scientists and engineers

    Dr. Rudolf KohulakDr. Rudolf KohulakLead AI ScientistPhD in applied mathematics, UCL. Masters in mathematical modelling.
    Dr. Ray OtsukiDr. Ray OtsukiLead ScientistPhD in theoretical physics, University of London. Six publications, including in the Journal of High Energy Physics. Earlier work in sell-side quantitative finance.
    Dr. Oliver SargentDr. Oliver SargentLead ScientistPhD in mathematics, University of Bristol. More than four years of postdoctoral research at the Weizmann Institute and the Technion. Built concept-based root-cause engines for a bank and an engineering company.
    Dr. Joey ReinessDr. Joey ReinessLead EngineerPhD in particle physics. MSci in physics, University of Cambridge.
    Dr. Edward BennettDr. Edward BennettAI ScientistPhD in mathematics. Masters in mathematics, University of Oxford.
    Dr. Euan MackayDr. Euan MackayScientistPhD in high-performance computational modelling, University of Dundee. MSc in theoretical physics, University of Edinburgh, on a Higgs Scholarship.
    Jonathan RubinJonathan RubinScientistPhD candidate in AI and machine learning, Imperial College London. Research in mechanistic interpretability and domain-specific language models.
    Avigail MarkevichAvigail MarkevichFull-stack EngineerBSc in computer science.

    What we will prove next

    Four things an outsider can check.

    1. A public releaseWeights, code and a technical report, under a permissive licence.
    2. A standing on the benchmarksResults on the public tabular benchmarks that anyone can rerun.
    3. A benchmark for readabilityThe first public test of how well a tabular model explains itself. Whoever sets the yardstick owns the category.
    4. Tier 1 referencesTwo or three large companies running Able V1 in production, with measured returns.

    We are building the readable tabular model, and looking for the partners to build it with.

    The field filled up in a single year. The model that sets the standard for readability will be decided in the next one, and it will not be decided slowly.

    Contact Erez Raanan, CEO

    • ResearchStaying at the frontier as these models scale up.
    • Industry versionsModel lines for telecom, banking and industry, built with the companies that run them.
    • ProductionThe enterprise product, and putting it to work at tier 1 companies.
    Mathlabs. Foundation models that run real systems.© 2026 Mathlabs