mLSTM Kernels
NXAImLSTM Kernels (Tiled Flash Linear Attention) provides fast, chunkwise-parallel Triton kernels for the mLSTM used in xLSTM, with JAX and PyTorch native implementations and Triton integrations for both. NXAI, the xLSTM team's company, distributes it under its own restrictive community license rather than an OSI-approved one.
Verified 2026-09-02 via the GitHub API and the LICENSE body.
Openness
2 high confidence- license
- NXAI-Community-License(non-OSI, Llama-3-style base with a revenue-bounded commercial term - a separate license is required above EUR 100,000,000 in consolidated annual revenue)
- source
- public
- core-gated
- ungated
LICENSE body is the NXAI Community License Agreement. Section 2, 'Additional Commercial Terms', requires a separate commercial license, granted at NXAI's sole discretion, once a licensee's consolidated annual revenue exceeds EUR 100,000,000 and it incorporates the material into a commercial product or service - a revenue-bounded commercial restriction, not an unconditional grant. The repository is public and unarchived and builds the whole product with no separate withheld component, so source is public and the core ungated - the license itself, not a gated core, is what caps the score.
- https://github.com/NX-AI/mlstm_kernels/blob/main/LICENSE recorded 2026-09-02
LICENSE body is the NXAI Community License Agreement. Section 2, 'Additional Commercial Terms', requires a separate commercial license, granted at NXAI's sole discretion, once a licensee's consolidated annual revenue exceeds EUR 100,000,000 and it incorporates the material into a commercial product or service - a revenue-bounded commercial restriction, not an unconditional grant.
- https://api.github.com/repos/NX-AI/mlstm_kernels recorded 2026-09-02
Repo metadata for mlstm-kernels's canonical repository - private false, archived false, fork false.
- https://github.com/NX-AI/mlstm_kernels/blob/main/README.md recorded 2026-09-02
README describes the library as providing fast, efficient mLSTM training and inference Triton kernels with JAX, PyTorch and Triton implementations, and cites the NXAI Community License Agreement governing the repository.
Adoption
2 high confidence11,616 PyPI downloads of `mlstm-kernels` in the trailing 30 days, which lands in the 10K-100K band of the software usage scale, level 2. `pip install mlstm_kernels` is the README's install path and the package is published by the NX-AI authors.
- https://pypistats.org/api/packages/mlstm-kernels/recent recorded 2026-09-02
last_month downloads = 11,616 for mlstm-kernels
- https://pypi.org/pypi/mlstm-kernels/json recorded 2026-09-02
PyPI metadata for mlstm-kernels 2.0.4 - summary 'A library providing fast and efficient mLSTM kernels for the xLSTM', authors Beck, Poeppel, Lippe and Boeck (nx-ai.com), license text NXAI COMMUNITY LICENSE AGREEMENT.
Capability
2 medium confidenceBanded on the category feature matrix as a narrow kernel set: kernels for one recurrent architecture family. Placed two bands below the tensorrt rung, the same level as flashkda and flashqla.
- https://github.com/NX-AI/mlstm_kernels/blob/main/README.md recorded 2026-09-02
README describes the library as providing fast, efficient mLSTM training and inference Triton kernels with JAX, PyTorch and Triton implementations, and cites the NXAI Community License Agreement governing the repository.
Verified 2026-09-02