April 2026 performance summary. Across four sgLang repositories, the team delivered scalable MoE capabilities, improved robustness, and performance optimizations that directly elevate production readiness and resource efficiency for large-expert models. Key features include larger-expert support in moe_align_block_size, LoRA MoE backend enhancements with Marlin integration, quantization configurability, Triton sgemm optimizations, and ongoing code quality improvements. Highlights: - Scalable MoE: Expanded moe_align_block_size to support more than 1024 experts with new CUDA kernels and Python bindings, enabling better utilization of multi-expert configurations. - LoRA MoE robustness: Decoupled LoRA MoE backend with Marlin support, virtual experts, and improved loading (logs, strict loading, embedding-name handling) for more reliable deployment. - Quantization control: Added a server CLI option to opt out of quantization (--quantization unquant) to give users explicit control over model behavior. - Backend performance: Optimized Triton-based sgemm and LoRA routing for faster inference and better batch/routing handling, with tests ensuring correctness. - Quality and consistency: Naming consistency improvements in CPU paths to reduce ambiguity and ease onboarding. Overall impact: Enhanced scalability for large-model deployments, improved reliability of LoRA MoE workflows, and greater control over performance characteristics, contributing to faster time-to-value for production ML tasks. Technologies/skills demonstrated: CUDA kernel development, Python bindings, LoRA MoE architecture, Marlin integration, virtual experts, robust model loading and logging, Triton sgemm optimization, routing logic, CLI configuration, and code quality improvements.