#99th of 23 in Open models
mamba
Reference PyTorch code for Mamba, Mamba-2 and Mamba-3 state space models
- Stars
- 18.9k
- License
- Apache-2.0
- Last commit
- Jul 2026
- Last release
- May 2026
- Language
- Python
Missed the latest GitHub check; these facts may be older.
Overview
Official implementation of the Mamba selective state space architecture, with Mamba, Mamba-2 and Mamba-3 blocks as PyTorch modules and CUDA kernels for the selective scan. It includes a sample language model backbone, generation and benchmark scripts, and pointers to pretrained 130M to 2.8B checkpoints on Hugging Face.
Who it is for: ML researchers and engineers building or evaluating state space models
Strengths
- Official code from the authors for Mamba, Mamba-2 and Mamba-3 blocks
- Pretrained checkpoints from 130M to 2.8B, auto-downloaded by the generation script
- Includes lm-evaluation-harness commands to reproduce paper zero-shot results
- Drop-in Mamba, Mamba2 and Mamba3 modules for your own PyTorch networks
Weaknesses
- Linux only; GPU use needs NVIDIA with CUDA 11.6+, AMD needs extra setup
- Pretrained models are base models only, no instruction tuning
- CUDA selective scan is opt-in and may compile locally with --no-build-isolation
- Research code: no model server or chat interface included
What it needs
- GPU required
- Needs PyTorch 1.12+, Python 3.10+, CUDA 11.6+, causal-conv1d (optional), lm-evaluation-harness (for evals)
- Models: Mamba, Mamba-2, Mamba-3, Hugging Face checkpoints (state-spaces)
Also in Open models
See all 23| Rank | Project | Score |
|---|---|---|
| 1 | llmfitRust tool that checks which LLMs fit your RAM, CPU and GPU | 86 out of 100 |
| 2 | layaSingle-pass multilingual classifier for typed choice, score and yes/no decisions | 83 out of 100 |
| 3 | timesfmPretrained time-series foundation model for zero-shot forecasting from Google Research | 74 out of 100 |
| #4 | unlimited-ocrOCR model that parses long documents and multi-page PDFs in one pass | 69 out of 100 |
| #5 | speechPyTorch toolkit for training and deploying ASR, TTS and speech LLM models | 67 out of 100 |