Flyto MLX
Copyright 2026 panwudi and Flyto MLX contributors

This product is a derivative work of oMLX:
  https://github.com/jundot/omlx
  Copyright (c) jundot and the oMLX contributors
  Licensed under the Apache License, Version 2.0

oMLX itself adapts code from vllm-mlx, mlx-lm, mlx-vlm, and mlx-audio,
each of which retains its own copyright and license. See LICENSE for
the full text of the Apache 2.0 license that governs this work.

Substantive additions on top of upstream oMLX as of this release:

  * /v1/chat/completions support for OpenAI input_audio content parts,
    wiring Gemma 4 nano (E2B / E4B) audio tower end-to-end through the
    chat endpoint.

  * DFlash Path A double-engine implementation, generalised across Qwen
    and Gemma 4 backends with optimised drafter co-loading.

  * macOS 26 (Tahoe) NSStatusItem occlusion-bit fix for the menubar GUI
    bundle.

  * Backports of five upstream-fixed-but-not-yet-released bug fixes:
    lm_head tokenizer vocab_size source, TokenBuffer seeding on cached
    prompt hits, health-check Session reuse, and others.

  * Chinese model aliases / presets for Qwen 3.5, DeepSeek V4, Gemma 4
    families.

  * Gitee mirror and ModelScope model registry support for users in
    mainland China.

For attribution requirements under Apache 2.0 §4, redistributors must
retain this NOTICE file along with the LICENSE file in any distribution
of Flyto MLX or any derivative thereof.
