Release Notes

For a detailed overview of what’s new in version 10 of Numerics.NET, see the dedicated page What’s New in Version 10.0.

This page lists the improvements since the original release of version 10 in November 2025.

Version 10.3.0

February 2026

Introduces a new random number generation stack designed for numerics workloads, centered on RandomSources in the Numerics.NET namespace.
Provides a consistent set of random primitive operations (integers, floating-point, bytes) and bulk generation APIs for high-throughput use.
Establishes cross-platform parity for seeded generation across Windows, Linux, and macOS, and across x64 and ARM64, within a major version (given the same call sequence).
Supports state snapshot and restore for checkpointing, exact continuation, and durable experiment reproduction.

Adds first-class support for independent random number streams suitable for parallel and concurrent code.
Enables deterministic branching from a base seed using stream identities (Stream IDs), so each worker or task index maps to a stable stream rather than “the next stream in line.”
Helps eliminate parallel flakiness by decoupling results from scheduling and execution order, assuming stable stream usage.
Includes guidance and APIs for common parallel patterns, from simple worker pools to more advanced stream management.

Adds a unified seeding and initialization model via SeedProfile, including opt-in compatibility with supported external ecosystems such as NumPy and C++.
Expands small seeds into full generator state with robust mixing, reducing the risk of correlated streams that can produce misleading results.
Provides full bidirectional interoperability with System.Random so modern generators can be used with legacy APIs, and existing System.Random instances can be used with Numerics.NET APIs.
Notes interop limitations and best practices (use adapters at API boundaries, prefer native sources in hot loops).

Greatly expands the set of available modern RNG algorithms under Numerics.NET.Random, covering general-purpose, stream-friendly, and parallel-oriented generators.
Updates Numerics.NET components to accept random sources directly, while retaining System.Random support for compatibility (including areas such as global optimization).

January 2026

Added support for persisting statistical models to JSON for deployment and reuse.
Persisted models store only the minimal predictive state, including:
- Model coefficients and parameters
- Transformation matrices (e.g., PCA loadings)
- Cluster centers for clustering models
- Category labels and schema information
Training data, diagnostics, residuals, and optimization history are not stored, reducing size and protecting data privacy.
The persistence format is versioned and backward compatible.

Introduced the Deployed model state for lightweight inference.
Deployed models:
- Provide the same Predict and Transform APIs as fitted models
- Use significantly less memory
- Can be used immediately after loading from JSON
Variable mappings and categorical encodings are preserved for correct predictions.

New ListVector<T> type for representing variable-length (ragged) data.
Supports per-list aggregation integrated with the existing grouping and aggregator infrastructure.
Enables elementwise transforms across lists, including in-place updates with MutableValues semantics and span-based access.
Provides factory methods on Vector for creating list vectors from nested data, flattened vectors with groupings, or matrix rows.
Supports flattening back to a standard vector and conversion to fixed-width row matrices with padding and truncation control.
Serves as the underlying representation for list-valued columns in data frames.

December 2025

New simple factories (Ones, full) for objects with constant valueds.
Shape-preserving “like” constructors for generating objects matching existing dimensions.
Support for random vector and matrix creation using arbitrary probability distributions.
Dense generators for structured matrices, including Toeplitz and Hankel matrices.

SignalMath now includes a full set of analytical signal generators for test signals, waveforms, pulses, chirps, and noise.
Generators operate directly on numeric sequences with optional span-based or allocating overloads.
Consistent function signatures support implicit discrete coordinates or explicit x-values.

November 2025

Added support for .NET 10 and corresponding C# 14 features.
Dropped support for .NET Core 3.1 and .NET 6.0. Both platforms have reached end-of-life and are no longer supported by Microsoft. Applications targeting these runtimes should migrate to .NET 8 (LTS) or later to continue receiving updates and security fixes.
All binary serialization APIs have been removed.

New unified factory API for vectors and matrices (CopyFrom, Wrap, Zeros, Identity, FromRows, etc.) with clearer semantics.
Compound assignment operators for vectors and matrices (.NET 10 only).
Native libraries have been updated to the latest stable release (Intel OneAPI 2025.3).

New static Interpolation class with linear, cubic, Akima, and inverse-linear methods.
New 1D interpolation methods: log-linear, log-log, periodic, circular.
New SignalMath convenience API for FFT, convolution/correlation, windowing, and PSD estimation.
New optimized convolution/cross-correlation kernels with configurable padding and kernel anchor.
New built-in nonlinear curve types, including: Michaelis–Menten, Gompertz, Richards, Weibull CDF, EMG, Pseudo-Voigt, Damped Sine, Power, and Double Sigmoid.

Want to go further back? See release notes for version 9.