OptiRS 0.3.0 Released — Pure Rust ML Optimization Suite Powered by SciRS2

The machine learning optimization foundation of the COOLJAPAN ecosystem just reached full production readiness.

Today we released OptiRS 0.3.0 — a comprehensive, production-grade pure Rust optimization library built exclusively as an extension of SciRS2-Core.

No Python. No PyTorch optimizers. No external crates.
No direct ndarray/rand usage (forbidden by design).
Just clean, memory-safe, hardware-accelerated optimizers that compile to a single static binary (or WASM) and run everywhere — from laptops to browsers to edge GPUs to cloud clusters.

Why OptiRS 0.3.0 is a game changer

For years, ML training meant depending on Python frameworks (PyTorch, TensorFlow) or fragile C++/CUDA optimizers.

These tools are powerful but suffer from:

• Python interpreter overhead and GIL bottlenecks
• Memory unsafety in custom CUDA kernels
• Vendor lock-in and complex dependency chains
• Poor WASM/embedded/no_std support
• Difficulty in reproducible, production-grade optimization research

OptiRS 0.3.0 ends all of that.

It delivers massive speedups while staying 100% within the SciRS2 ecosystem.
Notable results:

• SIMD-accelerated first-order methods: 2–4× faster
• ParallelOptimizer: 4–8× on multi-core CPUs
• GPU backends (CUDA/Metal/WebGPU): 10–50× vs CPU baseline

Technical Deep Dive: How We Built a Production-Grade ML Optimizer in Pure Rust

OptiRS was deliberately separated from SciRS2 to enable focused development, independent releases, and specialized hardware acceleration — while requiring full use of scirs2-core for every operation.

The architecture is clean and layered:

1. Core Optimizers (optirs-core)
   22 production-ready optimizers:

   • First-order (20): SGD, SimdSGD, Adam, AdamW, RMSprop, Adagrad, AdaDelta, AdaBound, LAMB, LARS, Lion, Lookahead, RAdam, Ranger, SAM, SparseAdam, GroupedAdam, FedProx, ReptileOptimizer, MetaSGD
   • Second-order (2): L-BFGS, K-FAC, Newton-CG

2. Learning Rate & Analysis Tools
   7 schedulers (CosineAnnealing, OneCycleLR, etc.) + Gradient Flow & Loss Landscape analysis.

3. Hardware Acceleration

   • SIMD via scirs2_core::simd_ops
   • Parallel via work-stealing scheduler
   • GPU/TPU via scirs2_core::gpu (multi-backend)

4. Advanced Modules (alpha → stable in 0.3.0)
   optirs-learned (Transformer/LSTM meta-optimizers), optirs-nas (evolutionary/RL/DARTS), optirs-bench (Criterion.rs statistical benchmarking).

Key Rust advantages:

• Strict dependency rule: only SciRS2-Core allowed (no external crates)
• Zero-cost abstractions + SIMD everywhere
• Full no_std + alloc readiness
• Memory-efficient gradient accumulation & chunking

What’s inside 0.3.0 (released March 18)

• Full production readiness of optirs-core
• All 22 optimizers stabilized with comprehensive monitoring & metrics
• SIMD, parallel, and GPU acceleration fully integrated
• Expanded learning-rate schedulers and analysis tools
• Production readiness confirmed with 1,220 unit tests + 76 doctests across 7 crates
• 251,000+ lines of pure Rust — zero Clippy warnings, fail0 enforced

This is the foundation

OptiRS is now the official optimization backend for the entire COOLJAPAN scientific and ML stack (total ecosystem: 21M SLoC Rust, 597 crates, 40+ production-grade libraries):

• SciRS2 / NumRS2 — all core training loops and metrics
• OxiBLAS — accelerated gradient math
• OxiMedia — vision model training pipelines
• Spintronics / OxiHuman — physics-informed and biomechanical optimization
• ToRSh / OxiRAG — high-throughput RAG and tensor training
• Future integration with OxiLean for formally verified optimizers

Repository: https://github.com/cool-japan/optirs

Star the repo if you want production-grade ML optimization without Python or external dependencies.

The era of “just use torch.optim” with all its overhead is over.

Pure Rust ML optimization — fully integrated with SciRS2 — is here, fast, safe, and sovereign.

— KitaSan at COOLJAPAN OÜ March 18, 2026