We are happy to announce the official release of Walnutpie version 0.0.1. Walnutpie is an MCMC sampler for continuously differentiable densities coded in Python, accepting models coded in Stan, PyMC, NumPyro, JAX, and plain old Python. Walnutpie is not an official Stan project I thought this was w…

Cool, looking forward to trying it out. I’ve already run walnuts on a few models with bridgestan, but that was a bit tedious to configure and run each time, so I’m very glad to see this interface. [image] Bob_Carpenter: Walnutpie is not an official Stan project I thought this was worth saying u…

[image] jonah: I know @WardBrian had expressed interest in this in the past. Brian, is that still something you’re interested in? It is! First I’d like to build the equivalent of what we have in Python in R (at least for Stan models, maybe not arbitrary R functions) and monitor what people di…

I realized in a moment of panic after we did the release that we hadn’t tested on anything but a cherry-picked set of models for illustration purposes. So I went home last night and vibe-coded (Opus 4.8 hard) some posteriordb tests. Thankfully the AI could figure out the posteriordb interface in P…

Hi Bob, on the 8 schools model I’m confused why walnuts fails with the centered parameterization. Nawaf shared this view of the adaptive step size with walnuts to capture the funnel. What am I missing here? [image]

This is really cool work! I love the BridgeStan idea and how you use Walnutpie. For fun, I’ve added Walnuts sampling (without Rhat early stopping) to the Stanli website and made it so you can watch it sample live at the same time as NUTS / dynamic HMC in your own browser: stanli: Stan Language Inte…

[image] seantalts: I think the main bad thing is that right now when you make a model it copies the entire interpreter over. The one bad thing about the API I guess, there wasn’t a clear way to keep shared BridgeStan state. The ‘sidecar manifest’ is a clever workaround. An alternative to cons…

Nawaf was looking at a pure funnel. He reated a really great movie, but the forums will only let me upload static images. @WardBrian may have a web link. The pure funnel is actually easier than the mess resulting from a handful of data points, which is worse in both centered and non-centered than…

oh sick. Done feat: embed the BridgeStan manifest in the data argument by seantalts · Pull Request #70 · seantalts/stanli · GitHub merging now, thank you!! import stanli from walnutpie import walnuts_stan model = stanli.bridgestan_model(stan_file="model.stan", data={"N": 10, "y": [0.1] * 10}) chai…

Someone posted on Stanli that it let them see an issue with warmup / stuck chain with Lotka-Volterra in WALNUTS. Posting here in case you guys are interested. [image] There’s one chain that spends some time post-warmup stuck and then corrects, but another chain is initted pretty far from a good p…

Walnutpie version 0.0.1 release

Publicity

WardBrian August 15, 2026, 1:16am 12

Yeah, we’ve independently run into this, with some alternative ideas for how to work around it at Models where Stan outperforms Nutpie/Walnuts - #39 by aseyboldt

We haven’t decided on any specific implementation change yet

Topic		Replies	Views
Models where Stan outperforms Nutpie/Walnuts Algorithms	47	2070	July 21, 2026
Comparing Stan's adaptation phase to that of nuts-rs? Algorithms	20	2120	August 11, 2023
Debugging a "stuck" sampler on a complex-ish model Modeling fitting-issues	9	322	July 14, 2026
Sparse NUTS: preconditioning with sparse matrix operations General	22	868	May 6, 2026
New adaptive warmup proposal (looking for feedback)! Algorithms	50	4893	July 31, 2020

Walnutpie version 0.0.1 release

Related topics