v5.4.0 (16 September 2026)

• Updated to Eigen 5.0.1. (#3271, #3338)
• Stan's vendored dependencies are now included with -isystem by default. (#3333)
• Added integrate_1d_double_exponential and integrate_1d_gauss_kronrod. (#3326)
• Added laplace_latent_solve and laplace_latent_solve_tol to return the posterior mean and Cholesky factor from the embedded Laplace approximation. (#3337, #3345)
• Added reshaping helpers for converting between matrices and arrays of Eigen vectors/row vectors. (#3312)
• RNGs now behave properly when given a size-0 input. (#3389)
• softmax and log_softmax now support row_vector and arrays of either. (#3313)
• Added missing checks to a few unconstraining functions used by initialization. (#3387)
• Updated student_t_qf to only returns column eigen vectors. (#3317)
• Fixed an issue with reduce_sum to allow it to take in tuples and tuples of tuples. (#3376)
• Fixed an issue with rep_matrix sometimes returning a row-major matrix. (#3342)
• Fixed bernoulli_logit_lpmf and bernoulli_logit_glm_lpdf upper-tail gradients. (#3397)
• Fixed an incorrect derivative of wiener_lpdf with respect to the relative starting point parameter w. (#3324)
• Added map, mapN, row_map, col_map, row_mapN, and col_mapN functors for elementwise and rowwise/columnwise mapping over std::vector and Eigen matrices. (#3346)
• Fixed a bug in ordered_logistic_lpmf and ordered_logistic_glm_lpmf higher order autodiff. (#3327)
• log_softmax now returns empty vector on empty input. (#3328)
• Fixed missing #include for finite_diff_hessian_auto (#3354)
• Updated normal_lccdf and std_normal_lccdf to use normal_lcdf and std_normal_lcdf implementations. (#3363)
• Removed extra std::forward in fwd/functor/reduce_sum loop. (#3323)
• Used std::nth_element to reduce algorithmic complexity of quantile implementation, support using Eigen containers for multiple p. (#3358)
• Fixed typo in tuple_to_laplace_options forwarding function. (#3330)
• Fixed documentation signatures for laplace functions with built-in bernoulli_logit likelihood. (#3355)
• Fixed make_holder(...) with var_value<Matrix> output to now eagerly execute. (#3334)
• Added const coeffRef() definitions for the Holder class to resolve downstream compilation errors. (#3352)
• Updated the .val() member function to always returns a CwiseUnaryOp when called with a var type. (#3351)
• Efficiency: use multiplication over std::pow for squaring. (#3373) (#3375)
• Removed dependency on boost/optional and boost/lexical_cast. (#3378)
• Fixed reduce_sum_static to only evaluates inputs once. (#3384)
• Fixed an issue with forward-mode jacobian for mappings R^n -> R^m with m != n. (#3386)
• Fixed use-after-move in cholesky_decompose and columns_dot_product. (#3390)
• Fixed an include issue with the QoS header on older Macs. (#3402)