Open-source MMM in 2025: Robyn vs PyMC-Marketing vs LightweightMMM.

Marketing mix modelling went from boutique-consultant-only to open-source-and-in-house between 2022 and 2024. Three tools now cover 90% of the in-house MMM work we see: Meta's Robyn, the PyMC Labs community's PyMC-Marketing, and Google's LightweightMMM. All three are free. None are equivalent.

Robyn (R)

Strengths: Fastest to a first model. Nevergrad hyperparameter optimisation is mature. Ridge regression is the default, which makes coefficients interpretable and stable. Active Meta Marketing Science support and a large user community.

Weaknesses: R-based, which is a team skills issue. No probabilistic uncertainty intervals out of the box. Adstock and saturation transforms are flexible but can over-fit without strong priors.

Fit: DTC and consumer brands with a quarterly planning cadence and ~18 months of weekly data. Teams that want answers, not debates about Bayesian philosophy.

PyMC-Marketing (Python)

Strengths: Full Bayesian inference. Credible intervals on every coefficient. Lets you encode domain knowledge as priors, which matters when you have 40 channels and 80 weeks of data. Active development, backed by PyMC Labs.

Weaknesses: Slower to fit (hours, not minutes, for a large model). Needs someone who can read trace plots and debug divergences. The learning curve is significant if nobody on the team has done Bayesian modelling before.

Fit: Sophisticated in-house analytics teams. B2B companies where decision stakes are high and "our best guess is between $X and $Y with 80% confidence" is a better answer than a point estimate.

LightweightMMM (Python, JAX)

Strengths: Google-built, JAX-backed, genuinely fast once compiled. Solid defaults. Good documentation.

Weaknesses: Development appears to have slowed in 2024. Issue tracker responses are slower than Robyn or PyMC-Marketing. Less community content for troubleshooting.

Fit: Teams already on GCP and JAX who want speed, and are not relying on active upstream development.

The honest take

If you are starting in 2025, we would pick PyMC-Marketing for a sophisticated in-house team, Robyn for everyone else. LightweightMMM is fine but we would not start new projects on it without a commercial backer.

None of the three replaces incrementality testing. Use them together.

Sources: Robyn documentation; PyMC-Marketing; LightweightMMM repo; internal client implementation work, 2023-2024.