mamba2

Rust-native MoE inference runtime with custom CUDA kernels for Blackwell GPUs. Includes DFlash speculative decoding, multi-tier Engram memory, and entropy-adaptive routing. Targets Qwen3.5-35B-A3B on a single RTX 5060 Ti 16GB.

Hybrid Mamba-2 + Transformer 2.94B LLM (Nemotron-H style) — Korean 3B model pretrained from scratch on 7× NVIDIA B200 GPUs with SFT + DPO alignment

Hybrid SSM-Attention language model on Apple Silicon with MLX — interleaving Mamba-2 and Transformer for efficient inference

The fastest Mamba2 SSD kernel for TPU written in Pallas

UAE / Saudi LLM llamafile — Single-file, no-install executables for Falcon, Jais, ALLaM, and more.

Helix-v2 1B bilingual hybrid LLM - Nvidia

Research prototype for real Mamba-2-130M inference under CKKS/FHE (OpenFHE/FIDESlib-GPU): +0.12% PPL surrogate and a verified 24-layer, 3-token encrypted B300 path.

This is a complete testing and construction project for a recurrent small-parameter language model based on the Mamba2 architecture.这是一个完整的基于mamba2架构的循环小参数语言模型的测试与构建项目.And it try to be built with a Mano Optimiters.Mano is a new Optimister

[RWKV6/mamba2(parallel fusion)] SSM

VerDee: Vertical Deep Network on Mamba-2 — staged LoRA (shallow/mid/deep), early-exit routing, and domain experts. Pilot on 370M; 16GB GPU friendly. Research/experiment repo.

Build Your Own Mamba — From Math to Metal

A 390M-parameter Mamba2 + Differential Attention hybrid language model

GLACIER: Mamba with indexless memory. This project integrates the Mamba SSM with ICE-Lite, a virtual memory engine, to solve context rot. By adding persistent, time-aware memory, GLACIER gives Mamba the long-term recall of a Transformer while retaining its $O(N)$ speed. Apache 2.0 licensed, by Dopove.

Native AscendC Mamba2 selective scan / SSD forward-backward custom operator for Huawei Ascend 910B3 and 950PR, with CANN, torch_npu, A100 benchmarks and msprof profiling.

Unofficial PyTorch reproduction for Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality.

Run NVIDIA's Nemotron-Labs-TwoTower-30B-A3B block-wise diffusion LM natively on Apple Silicon with MLX — AR tower + full two-tower mask-diffusion, 4/6/8-bit + bf16.

llama.cpp fork with additional SOTA quants and improved performance

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

How much cheaper is it to inject a retrieved fact next to the answer than to make a Mamba-2 model recall it from context? A small, controlled study: inject vs long-context, 2-hop chains, and adjacency decay.

A research notebook on competing sequence modeling paradigms: implement, benchmark, and understand the architectural bets.