Release history

mixle releases

Published package releases under the mixle name, with notes copied into a concise changelog format and links back to the release artifacts.

Latest 0.7.0
Published July 9, 2026
Package files wheel and source distribution
Status Beta

Latest release

mixle 0.7.0

Generic AI-capability platform pieces on top of the core estimation engine: task decomposition, cross-modal reasoning, self-improvement loops, and a system facade, plus a hardening pass across the automatic-inference and design-of-experiments subsystems.

New model families

PINNRegression, HamiltonianNet, make_deep_set, monotonic MLPs and input-convex energy networks, CopulaDistribution, and GatedMixtureDistribution (input-dependent mixture-of-experts weights).

Task decomposition and agents

Plan models fit as Markov chains over agent traces, an outcome-trained decomposer, a minimal orchestrator loop, a local model registry, bit-identical execution replay, and the Receipt object (ledger + trace + calibration + provenance, offline re-verifiable).

Cross-modal reasoning

ModalityView, per-edge conditional-transport premise checks, belief walks across chains of verified transports, cycle-consistency as a self-supervised abstention signal, and task-sufficient projection.

Self-improvement loops

A shared collapse monitor across amplification loops, a composition operator, diagnosis-directed correction, a knowledge-accumulation flywheel measurement, and cost-aware routing threshold selection.

Release-verification hardening

A fresh, non-editable wheel install caught a broken import mixle (an eager-import cascade from one missing name), 8 modules importing torch-gated names unconditionally at module level, and a real DPOAccumulator.value() bug. All fixed; a full 687-module import sweep now gates every release from a bare install.

Release notes PyPI package Source tag

mixle 0.6.2

Focuses the 0.6 line around real task workflows: LLM teachers, local distilled models, neural density families, and safer artifacts without changing the core distribution/estimator API.

Task workflows

Added typed evidence surfaces for models that can retrieve context, compute, delegate to an LLM, cite sources, and abstain when confidence is low.

Artifacts and deployment

Added local-first telemetry, reproducibility checks, certificate helpers, placement planning, and job abstractions for production handoff.

Neural density families

Flow, MAF, VAE, and DiscreteAR can now be fit and composed inside model trees.

Distillation methods

Expanded task distillation with response training, dark knowledge, multi-teacher methods, hints, attention transfer, and sequence-level objectives.

Correctness fixes

Neural, DPO, and streaming Transformer accumulators now preserve sample weights; registry paths are constrained to the store root; structure-search cloning uses deep copies.

Release notes PyPI package Source tag What is new

Previous release

mixle 0.6.1

A feature release focused on neural distributions, learned dependency structure, cross-modal reasoning, task distillation, and inference tooling.

Neural densities

Added composable neural distribution families including normalizing flows, masked-autoregressive flows, VAEs, mixture density networks, autoregressive categorical densities, conditional flows, and energy-based models.

Dependency learning

optimize(data) and fit(data) can discover cross-field dependency graphs by default, with BIC-gated structure selection.

Cross-modal reasoning

Added product-of-experts and hybrid attention/structured fusion, transfer-safe adapters over frozen encoders, iterated-EKF assimilation, and retrieval by model belief.

Task distillation

Expanded mixle.task with solve() support for regression, multi-label, and structured routines, plus conformal calibration, generative-text students, agent-trace harvesting, plan generation, and traffic health monitoring.

Inference and tooling

Added closed-form Fisher-weighted parameter merge, variational projections, exact GMM collapse, Runnalls reduction, multi-prime NTT convolution for count-DP ranking, and documentation updates.

Release notes PyPI package Source tag

First mixle release

mixle 0.6.0

First release under the mixle package name, replacing the earlier pysparkplug / pysp naming.

Package rename

Published as mixle on PyPI with import mixle as the public import path.

Automatic inference

Models and estimators share the same shape, letting fit choose conjugate, EM, MAP, variational, or MCMC routes from the structure being fit.

PPL language core

Added deterministic-expression slots, potential() custom factors, hierarchical .each() forms, data-indexed latents, non-Normal GLMMs, and per-parameter R-hat / ESS summaries.

Fitting API

optimize and fit accept distribution prototypes or data directly, infer estimators when possible, and forward the same fitting controls.

Streaming estimators

Added free-dimension inference for categorical and Dirichlet forms, alongside unified streaming-estimator behavior.

Release notes PyPI package Source tag

Sources

Source references.

Dates and package artifacts are checked against PyPI. Narrative notes are summarized from GitHub Releases. If future releases add more detail, this page should be updated from those two sources first.

PyPI release history Published package versions and upload artifacts. GitHub Releases Release notes, tags, and source archives. GitHub tags Version tags when comparing source snapshots.