Rapid-MLX
Copyright 2025-2026 the Rapid-MLX contributors

This product includes software developed by third parties. The notices below
are provided under section 4 of the Apache License, Version 2.0, and under the
terms of the other licenses named here.


Origin of this codebase
=======================

Rapid-MLX began as vLLM-MLX (`vllm-mlx`), created by Wayner Barrios and
published at:

    https://github.com/waybarrios/vllm-mlx

licensed under the Apache License, Version 2.0. The initial commit of this
repository (2025-12-06) is that project, and development continued in this
history until 2026-02-15. The project was renamed to Rapid-MLX on 2026-03-13.

**The files in this repository have been changed.** Since the rename the
codebase has been extensively modified, extended, and in large part rewritten
by the Rapid-MLX contributors; the Python package now uses the `rapid_mlx`
import name and retains `vllm_mlx` as a deprecated compatibility shim. Neither
the original authors nor any upstream project endorses this distribution. The
complete record of what
changed, when, and by whom is preserved in the git history of this repository.

Rapid-MLX is not affiliated with, and is not a distribution of, the vLLM
project (https://github.com/vllm-project/vllm).


Vendored third-party source
===========================

The Bonsai 2 Hadamard runtime in
`rapid_mlx/models/prism_hadamard_qwen35.py` is adapted from Prism ML's
https://huggingface.co/prism-ml/Ternary-Bonsai-2-27B-mlx-2bit
at revision 3f926b415992eaa2ae9dd7b573706494d6bbf787. The runtime directory
carries the MIT License — Copyright © 2023 Apple Inc. — reproduced in the
adapted file. Created using Bonsai by Prism ML; the model's root NOTICE also
credits Prism ML, Inc. (2026-present) and Alibaba Cloud under Apache-2.0.
Model weights remain separate; Python from model snapshots is never executed
by this loader.

The Qwen4/Qwen3.8 stage-one QSA selector in
`rapid_mlx/kernels/qsa_stage1.py` is adapted from
https://github.com/pierre427/mlx-lm-unified at commit
685dff6c1ee602f1bae2e51c758a3b07f593a0c2. Its radix-selection design is
adapted from MTPLX PR #397. The file is licensed under the MIT License;
see `rapid_mlx/kernels/LICENSE-QSA-STAGE1`.

The following directories contain third-party source redistributed inside this
repository. Each keeps its upstream license text alongside the code.

* rapid_mlx/compiled_decode.py and rapid_mlx/compiled_precision.py
  Shape-stable cache/replay and precision-routing design adapted under the
  MIT License — Copyright © 2023 Apple Inc.
  See rapid_mlx/COMPILED_DECODE_LICENSE.

* videox_fun_mlx/
  CogVideoX-Fun MLX runtime. Vendored from
  https://github.com/dgrauet/VideoX-Fun-mlx at commit
  26326e7d52e6762375227b320d77003dac764d14 (2026-07-21), which is a fork of
  the original VideoX-Fun by the Alibaba PAI team:
      https://github.com/aigc-apps/VideoX-Fun
  Licensed under the Apache License, Version 2.0.
  See videox_fun_mlx/LICENSE and videox_fun_mlx/NOTICE.

* rapid_mlx/audio/sa3/
  MLX Stable Audio 3. Vendored from
  https://github.com/Stability-AI/stable-audio-3 at commit
  124e8a799f57a1f665495ecb72e547d0a62867f1.
  MIT License — Copyright (c) 2026 Stability AI.
  See rapid_mlx/audio/sa3/LICENSE and rapid_mlx/audio/sa3/NOTICE.

* rapid_mlx/image/hidream_runtime/
  HiDream-O1-Image-Dev MLX text-to-image runtime adapted from
  mlx-community/HiDream-O1-Image-Dev-mlx-bf16 at commit
  33c7a00bce8e3410304f83ec408a15a1eb6782df.
  MIT License — Copyright (c) 2026 mrbizarro and contributors.
  See rapid_mlx/image/hidream_runtime/LICENSE and NOTICE.

* rapid_mlx/image/sdxl_runtime/
  Stable Diffusion XL MLX text-to-image runtime adapted from
  https://github.com/amirhossein-razlighi/mlx_diffuser at commit
  a26b42aee4e31999dbb4429226b66d896d49e1d8.
  CC0 1.0 Universal — original work by Amirhossein Razlighi and contributors.
  See rapid_mlx/image/sdxl_runtime/LICENSE and NOTICE.

* rapid_mlx/image/sd35_runtime/
  Stable Diffusion 3.5 Large MLX runtime adapted from DiffusionKit at commit
  498e5dba5fb48b0f01cb6b5c2292a6dbea67a317.
  MIT License — Copyright (c) 2024 Argmax, Inc.
  See rapid_mlx/image/sd35_runtime/LICENSE and NOTICE.

* apps/rapid-mac/Vendor/SwiftMath/
  Vendored from https://github.com/mgriebling/SwiftMath at 1.7.3
  (fa8244ed032f4a1ade4cb0571bf87d2f1a9fd2d7) with a resource-resolution patch.
  MIT License — Copyright (c) 2023 Computer Inspirations.
  The bundled math fonts carry their own LICENSE, OFL.txt, and
  GUST-FONT-LICENSE.txt notices alongside the font files.
  See apps/rapid-mac/Vendor/SwiftMath/LICENSE.

* rapid_mlx/patches/glm5_next_processor.py
  GLM-5 Next torch-free processor adapted from
  https://github.com/jundot/omlx at commit
  c520d7e6124cf6255868e589c6beb75cf4cee8f9.
  Licensed under the Apache License, Version 2.0.

* rapid_mlx/patches/glm5_next_runtime.py
  GLM-5 Next text-runtime compatibility logic adapted from the mlx-vlm
  correctness work at commits f9e2c507a154575956ea97449729a67de72c9c70 and
  bffd4856a4504b76a242536f2dd356093c8e2e92.
  Licensed under the MIT License.

* rapid_mlx/speculative/native_mtp/transaction.py and glm5_compat.py
  Cache-owned MTP transaction and stateless GLM drafter adapter adapted from
  the mlx-vlm implementation in
  https://github.com/Blaizzy/mlx-vlm/pull/2206.
  Licensed under the MIT License.

* rapid_mlx/models/deepseek_v41_native/
  DeepSeek V4.1 native MLX runtime. The target runtime is adapted under the
  Apache License 2.0; the DSpark draft algorithm follows the model publisher's
  MIT-licensed reference implementation.
  See rapid_mlx/models/deepseek_v41_native/LICENSE,
  rapid_mlx/models/deepseek_v41_native/LICENSE-DSPARK, and NOTICE.


Runtime dependencies
====================

Rapid-MLX is built on Apple's MLX and the MLX model runtimes. These are
declared dependencies installed from their own distributions, not vendored
here, and each is used under its own license:

    MLX          https://github.com/ml-explore/mlx            MIT
    mlx-lm       https://github.com/ml-explore/mlx-lm         MIT
    mlx-vlm      https://github.com/Blaizzy/mlx-vlm           MIT
    mlx-audio    https://github.com/Blaizzy/mlx-audio         MIT

The macOS desktop application enumerates the components it ships in
apps/rapid-mac/THIRD_PARTY.md, and reproduces their license texts inside the
application bundle.

Optional Python extras pull in additional dependencies under their own terms,
including copyleft licenses. Those extras are not installed into the shipped
desktop application. See apps/rapid-mac/THIRD_PARTY.md for the scope of what
the binary contains.


Model weights
=============

No model weights are distributed with this software. Weights are downloaded at
runtime from their publishers and are covered by their own separate licenses,
which are not granted by this NOTICE or by the accompanying LICENSE.
