jaxglobal and discopt

discopt is the author's other MINLP solver, and the more mature of the two. They solve the same class of problems with different designs. This list compares them as of October 2026; the discopt facts come from discopt's own README and development notes.

	jaxglobal	discopt
Model input	a plain JAX function, traced to a jaxpr, or an .nl file. jnp.where logic, neural networks, GP surrogates and fixed-step ODE integrators (Python loops or lax.scan) get certified bounds.	an algebraic modeling API (Model, m.continuous, dm.exp, ...), or .nl, .gms and QPLIB files. Opaque Python callables (dm.custom, dm.external) are solved without a dual bound.
Implementation	pure Python and JAX: relaxations, the LP interior-point method, FBBT and OBBT are all jit/vmap kernels.	a Rust core (expression IR, B&B tree, .nl parser, FBBT, simplex) with a numpy relaxation layer and Python orchestration. JAX is no longer on its default solve path.
Lower bounds	McCormick and alphaBB rules applied to each jaxpr primitive; alphaBB is derived automatically from grad(grad(prim)). Also a lifted quadratic and polynomial LP with RLT rows, and convex outer approximation.	McCormick over 28 primitives with a table of univariate envelopes, piecewise McCormick, alphaBB, G-convexity, RLT, PSD/SOC, GMI and OA cuts.
Node LPs	a primal–dual interior-point method in JAX (Mehrotra predictor–corrector, early stop), batched over boxes, with weak-duality-certified bounds.	an in-house Rust simplex with sparse LU and warm starts; HiGHS only for pure LP/MILP.
Search	batched spatial B&B: many boxes bounded per vmap call, pseudocost or width branching, and symmetry pruning on integer polynomial problems.	a native Rust spatial B&B, plus AMP (adaptive multivariate partitioning: piecewise MILP relaxations iterated against an NLP) and a signomial global optimizer.
Primal heuristics	multi-start local NLP (pounce when installed, otherwise optax L-BFGS; scipy on request), dives, rounding repair and integer polish.	multi-start NLP, feasibility pump, diving, RINS, local branching and QUBO local search.
Soundness	float64 re-check of every point; optional safe=True outward rounding; MINLPLib wrong-answer checks and a fuzzer.	outward rounding, MINLPLib .solu checks and certificate checks.
Fixed cost	JAX compilation: seconds per new model structure. A warm cache removes most of it.	small: native code with no compile step per model.
global50, 60 s	48 of the 50 run, 45 of those solved, 0 wrong (this repo, 58bb551, October 2026, before PR #76).	48 of 50 "ok" (incumbent matches the known optimum), at a geomean time about 8x BARON's (discopt report, July 2026).

What jaxglobal lacks from discopt. It has no AMP or piecewise MILP relaxations, no simplex warm starts, and no GMI or PSD cuts. Its heuristics are narrower, and every new model structure pays a compile cost. discopt is the better choice for problems written as equations.

What jaxglobal is for. It takes models that are only available as JAX code: surrogates, where logic and simulation loops. It bounds them by tracing, without rewriting them as equations, and its bounding kernels batch over boxes. B&B still runs on the host, and GPU throughput has not been measured (see Limitations).

Status. Development is paused at this point. None of the discopt features above are planned for jaxglobal yet.