mamba-3

Native Blackwell (sm_100) tcgen05 training backward for the gated-linear-recurrence family (GDN/KDA/GLA/SSD/LA), plus a contract-grade verifier that audits LLM-generated GPU kernels. Six Mamba-3 Triton kernels; GRPO trainer graded on contract correctness. arXiv:2608.12700

Native MLX reimplementation of the Mamba-3 architecture from state-spaces/mamba, adapted for Apple Silicon (MLX). The implementation mirrors the block-selection logic, core Mamba-3 recurrence concepts, inference-cache semantics, and config structure, while adding MLX-native generation, benchmarking, serialization, and LoRA fine-tuning. Open Source.

[NeurIPS'2026] Position Bias in Mamba and Hybrid Language Models | Evidence Position Bias Across Sequence Mixers in Long-Context Question Answering

Mamba-3 in a single file of pure PyTorch. Official mamba_ssm API, nothing to build, runs wherever PyTorch runs.

A simple, minimalistic, and explainable JAX implementation of Mamba 2 & Mamba 3

Reproducible benchmarks for state-space, attention, and hybrid sequence models on NVIDIA GPUs. Measures latency, memory bandwidth, compute utilization, and energy per token to quantify the GPU-SSM mismatch.

Dynamic State Space Ensemble for Multi Horizon Market Forecasting