For most of marketing mix modeling's modern life, the first real decision was not statistical. It was financial. A brand that wanted MMM picked up the phone to a consultancy, and the answer came back as a contract. Managed MMM engagements still run from roughly 50,000 to over 200,000 dollars a year depending on scope, and at the enterprise end six figures is the floor, not the ceiling. That price tag did one quiet thing for a decade: it decided who got to measure media scientifically, and a team that could not absorb the fee was simply locked out.
Open-source frameworks removed that gate. The modeling code that used to sit inside a vendor's black box is now on GitHub, free to install, with three serious options to choose from. What follows is a comparison of those three, written to explain how they think rather than to crown one winner, because the right answer depends entirely on the team holding the keyboard.
Why open-source MMM is a real change, and where it stops
The honest version of the open-source story has two halves, and most write-ups only tell the first.
The first half is true. A capability that cost a six-figure annual engagement is now a Python library. A data team can install it, fit a model against its own spend and sales data, and inspect every assumption the model makes, because the source is right there. There is no vendor methodology to take on faith. For a company that already employs analysts, the barrier that mattered most has genuinely fallen.
The second half is the part the marketing skips. Open-source frameworks lower the software cost and nothing else. They are model-building libraries, not finished products. They ship without data connectors, without a data preparation layer, and in two of the three cases without a usable interface for a non-technical stakeholder. The unglamorous work is unchanged: someone still has to build and maintain the pipeline that feeds the model clean weekly data, and the standing line among practitioners is that the model is about 20 percent of the work while consolidating the data is the other 80 percent. One analysis puts a first validated model at six to sixteen weeks from kickoff to production, with maintenance after. None of these tools is a button; each assumes a trained operator. Without a data scientist who is comfortable in Python and can read a model's diagnostics, the free licence changes very little. That is the threshold to weigh before any framework comparison. For whether your budget can support MMM at all, see our piece on the MMM budget threshold.
Google Meridian: the opinionated newcomer
Meridian is Google's open-source MMM framework. Google began testing it with brands in 2024 and opened it to everyone in January 2025, and of the three it is the most actively developed by a wide margin. The public repository shows dozens of commits in a typical month and a steady release cadence, the signature of a tool with a team behind it rather than a project in maintenance.
Three design choices define it. First, it is fully Bayesian. Meridian is built on TensorFlow Probability and samples with the No U Turn Sampler, so every result arrives as a distribution rather than a single number. Second, it is geo-level by design. Meridian's natural unit is a region, and it fits a hierarchical model across geographies, which gives it more data points to learn from and lets it borrow strength between regions instead of treating each one as an island. It still supports national-level modeling, but geo is where it is built to live. Third, it is opinionated. Meridian ships with explicit support for reach and frequency data, so it can model how often an audience saw an ad rather than only how much was spent, and it includes a Google Query Volume control to reduce a known bias in paid search measurement. Those are points of view baked into the framework. They make Meridian less of a blank canvas and more of a considered default.
The opinionated stance extends to usability, where Meridian has moved fastest. In February 2026 Google launched a Scenario Planner, a no-code interface built on Meridian that lets a marketer test budget allocations and read ROI estimates in a Looker Studio dashboard. A model whose output only a data scientist can open is a model most of the organization cannot use, and Meridian is the only one of the three with a serious answer to that.
The caveats are real. Meridian still expects a strong data science background to set up and run well, and it carries a clear Google flavor, from the Query Volume control to its fit with the Google ecosystem. The source is fully open, so this is not a lock-in trap, but a team should know it is adopting a worldview, not just a solver.
PyMC-Marketing: the framework for teams that want control
PyMC-Marketing comes from PyMC Labs, the consultancy behind the PyMC probabilistic programming library, and it is the other actively maintained open-source MMM. Like Meridian it is fully Bayesian and under active development with frequent commits. The difference is philosophy. Where Meridian hands you a considered default, PyMC-Marketing hands you a toolbox.
Flexibility is the entire point of it. Built directly on PyMC, the framework does not limit a team to the model it ships with. According to the project's own framework comparison, it supports custom priors and custom model terms, time-varying coefficients, hierarchical geographic modeling, and out-of-sample forecasting, and it integrates with experiment tracking through MLflow. If an analyst wants to change the shape of an adstock curve, encode a specific business belief as a prior, or add a component no MMM library anticipated, PyMC-Marketing allows it. It is also broader than its name suggests: the same package ships customer lifetime value and other marketing models alongside the MMM module.
That power has a cost, and the framework's own documentation is candid about it. PyMC-Marketing has no native user interface and no built-in budget optimizer or data connectors of the kind Meridian's Scenario Planner now provides. It is the choice for a team that wants to build its own MMM from the ground up and has the data science depth to do it. In the right hands that is a feature: a team with a real statistician will often prefer a framework that gets out of the way over one that makes decisions for them. In the wrong hands, the same flexibility is enough rope to build a confident, wrong model with no opinionated default to catch the mistake.
Meta Robyn: the framework that proved it could be done
Robyn deserves real credit, and it also deserves an accurate description of where it stands.
The credit first. When Meta released Robyn in 2020, open-source MMM as a category barely existed. Robyn did more than any other single project to change that. It took a technique associated with economics PhDs and six-month consulting timelines and made it something a competent analyst could run in code. Meridian and PyMC-Marketing are both, in part, answers to the demand it created.
Robyn also works differently from the other two, which is the most instructive thing about it. Robyn is not Bayesian. It uses ridge regression, a regularized form of linear regression that handles correlated channels by penalizing large coefficients, and it pairs that with an automated search over the model's hyperparameters using Nevergrad, Meta's evolutionary optimization library. Instead of asking an analyst to reason about priors, Robyn runs a multi-objective search across thousands of candidate models and balances statistical fit against business plausibility and calibration to experiments. It is a different answer to the same problem: automate the hard choices rather than ask the modeler to make them explicitly.
Now the accurate part. Robyn is no longer actively developed, and the evidence is not subtle. Its last tagged release was version 3.12.0 in December 2024. Its public repository shows only a handful of code commits across the whole of the last year, and recent issues filed by users sit open with no maintainer response. Compare that with Meridian's dozens of commits a month and the contrast tells the story without anyone needing to announce anything. Robyn is not deleted and the existing code still runs, but it has clearly entered a quiet wind-down. Anyone starting a new MMM program on Robyn in 2026 is building on a foundation that is no longer moving.
What does the wind-down signal? Partly simple consolidation: a young category is settling toward the two entrants with the most momentum. But there is a more pointed reading. Robyn was built by Meta, a company that sells advertising, and an MMM decides where the next advertising dollar goes. A platform building the measurement layer for its own ad inventory is a structural conflict worth naming, and it applies to Google's Meridian exactly as much as it applied to Meta's Robyn. The lesson is not that platform-built tools are dishonest. It is that open-source MMM matters precisely because the source is open: a team can read the model and check that the math is neutral. With a closed vendor model you trust. With an open one you verify, and that is the property to protect as the category consolidates.
Bayesian and frequentist, in plain language
The split between Robyn and the other two is the single most important concept for understanding open-source MMM, so it is worth stating without the jargon.
A frequentist model, which is what Robyn's ridge regression is, treats the answer it seeks as one fixed, unknown number. It searches the data for the single best estimate of that number and reports it: this channel returned 2.1 times its cost. Robyn wraps that in an automated search across many model configurations, but the output of any one model is still a point estimate.
A Bayesian model, which is what Meridian and PyMC-Marketing are, treats the answer as uncertain from the start and describes it as a range. It begins with a prior, an honest statement of what you believed before seeing this data, perhaps from a past experiment. It updates that belief with the data and reports the result as a credible interval: the return is most likely between 1.6 and 2.6 times, centered near 2.1. A 95 percent credible interval has the plain-English meaning people wrongly assume a confidence interval has, namely a 95 percent probability the true value sits in that range.
Why does this matter for a budget decision? Two reasons. First, honesty. A point estimate of 2.1 reads as certainty the model does not have. A credible interval tells a budget owner where the model is confident enough to act and where it is guessing, the difference between scaling a channel and testing it. Second, priors help when data is thin. Because a Bayesian model does not ask the data to prove everything unaided, it can produce sensible estimates with less history and tends to handle correlated channels more gracefully. None of this makes Bayesian automatically correct: a bad prior is a bad assumption, and a frequentist model with a clean search can be sound. But it explains why the actively developed frameworks are both Bayesian. For more on that shift, see our piece on the MMM renaissance.
A decision guide
Strip away the detail and the choice is mostly about the team.
Choose Meridian if you want an actively developed, opinionated framework with a clear default and the only credible no-code interface of the three. It suits a team that has data science capability but wants the framework to make sensible choices rather than expose every dial, and it fits geo-level modeling and advertisers who care about reach and frequency. You are adopting a worldview with a Google accent, which is fine as a conscious choice.
Choose PyMC-Marketing if your team has genuine statistical depth and wants maximum control. If your analysts want to specify their own priors, customize the model structure, and build an MMM that fits your business rather than the framework's assumptions, this is the one built for that. It asks more of the team and rewards them with fewer constraints.
Be cautious with Robyn for a new build. Its place in this category's history is secure, and studying how its automated search works is genuinely educational. But with development wound down, starting fresh on it in 2026 means building on a tool that has stopped moving while two well-supported alternatives improve.
And before any of that, the threshold that applies whichever framework wins. The licence is free. The data scientist, the clean weekly data, and the pipeline that keeps it flowing are not, and they decide whether your MMM is worth trusting. Open-source did not make marketing mix modeling cheap. It made it accessible to teams already equipped to run it, a smaller and more honest claim, and the right one to plan around.
Council summary
This post argues that open-source MMM changed who can do the work without changing how hard the work is: the licence fee fell to zero, but the need for a data scientist, clean weekly data, and a maintained pipeline did not. It teaches the three frameworks by design philosophy rather than ranking them: Meridian as the opinionated, actively developed default with the only credible no-code interface, PyMC-Marketing as the flexible toolbox for teams with real statistical depth, and Robyn as the pioneer whose development has clearly wound down. The Bayesian-versus-frequentist section is the conceptual core, explaining why both surviving frameworks report ranges rather than point estimates and why that honesty matters for a budget decision. The takeaway is a choice keyed to the team rather than the tool: open-source made MMM accessible to teams already equipped to run it, not cheap for everyone.
Comments