Swift-Qwen3.8-27B
Copyright 2026 UkisAI

UkisAI's contribution (the "Swift Contribution") is licensed under the
Swift Open License v1.0. See LICENSE.

This model is a Derivative Work of Qwen3.8-27B
  https://huggingface.co/Qwen/Qwen3.8-27B
  Copyright 2026 Alibaba Cloud
  Licensed under the Apache License, Version 2.0. See LICENSE-APACHE-2.0.

Changes made by UkisAI (Apache License 2.0, Section 4(b) change notice):
  - model-*.safetensors, model.safetensors.index.json: model weights were
    fine-tuned by UkisAI (LoRA adapter trained by UkisAI and merged into the
    Base Model weights).
  - generation_config.json: added "min_p": 0 and "repetition_penalty": 1.0.
  - README.md: replaced. ukisai-banner.png and swift-speed-demo.mp4 added.
  - All other files (config.json, chat_template.jinja, tokenizer.json,
    tokenizer_config.json, vocab.json, merges.txt, preprocessor_config.json,
    video_preprocessor_config.json) are unmodified from Qwen3.8-27B and
    remain under the Apache License, Version 2.0.

------------------------------------------------------------------------------
The notice above is reproduced unchanged from ukisai/Swift-Qwen3.8-27b.
What follows is an additional attribution and change notice for this
repository. It does not modify the Swift Open License v1.0 or the Apache
License, Version 2.0.
------------------------------------------------------------------------------

Swift-Qwen3.8-27B-PARO-MXFP6
  https://huggingface.co/hugypufy/Swift-Qwen3.8-27B-PARO-MXFP6

Copyright 2026 UkisAI. Swift Contribution licensed under the Swift Open
License v1.0 (https://huggingface.co/ukisai/Swift-Qwen3.8-27b/blob/main/LICENSE).
Derivative of Qwen3.8-27B, Copyright 2026 Alibaba Cloud, Apache License 2.0.

This repository is a quantized Derivative Work of Swift-Qwen3.8-27B
  https://huggingface.co/ukisai/Swift-Qwen3.8-27b
made by hugypufy. It is not made, reviewed or endorsed by UkisAI, by
Alibaba Cloud, or by z-lab.

It incorporates rotation tensors from z-lab/Qwen3.8-27B-PARO
  https://huggingface.co/z-lab/Qwen3.8-27B-PARO
  Licensed under the Apache License, Version 2.0. See LICENSE-APACHE-2.0.

Changes made by hugypufy, 2026-09-16 (Swift Open License v1.0 Section 4(b);
Apache License 2.0 Section 4(b)):
  - model-*.safetensors, model.safetensors.index.json: replaced with new
    weights (same 18-shard layout).
      * The 400 linear projections of the 64 decoder layers were rotated with
        the ParoQuant rotations of z-lab/Qwen3.8-27B-PARO and quantized to
        OCP MXFP6 E2M3 elements with e8m0 scales per 32 weights. The weights
        and block exponents were then fine-tuned layer by layer under that
        grid, with the rotations frozen, and exported as 6-bit codes.
      * The rotation tensors (theta, pairs, channel_scales) are copied
        unchanged from z-lab/Qwen3.8-27B-PARO.
      * All other floating-point tensors were converted from bfloat16 to
        float16.
      * The MTP head tensors (mtp.*) were removed.
  - config.json: "quantization_config" added. Otherwise unmodified.
  - README.md: replaced.
  - NOTICE: this addendum appended.
  - chat_template.jinja, generation_config.json, tokenizer.json,
    tokenizer_config.json, vocab.json, merges.txt, preprocessor_config.json and
    video_preprocessor_config.json are byte-identical to Swift-Qwen3.8-27B.
    LICENSE and LICENSE-APACHE-2.0 are unmodified copies.
  - ukisai-banner.png and swift-speed-demo.mp4 are not included.