Swift-Qwen3.8-27B-Uncensored-W4A16 and Swift-Qwen3.8-27B-Uncensored-W4A16-fast
Quantization and serving preparation: Copyright 2026 TyroneNel

This work incorporates the Swift Contribution, licensed under the Swift Open
License v1.0 (see LICENSE):

  Swift-Qwen3.8-27B, Copyright 2026 UkisAI.
  https://huggingface.co/ukisai/Swift-Qwen3.8-27b
  Derivative of Qwen3.8-27B, Copyright 2026 Alibaba Cloud, Apache License 2.0.

Provenance chain:

  1. Base Model: Qwen/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.

  2. Swift Contribution: ukisai/Swift-Qwen3.8-27b
     https://huggingface.co/ukisai/Swift-Qwen3.8-27b
     Copyright 2026 UkisAI. Licensed under the Swift Open License v1.0.
     See LICENSE. A reasoning-efficiency LoRA adapter merged into the Base
     Model weights (see UkisAI's NOTICE and model card).

  3. Uncensoring: d0xin/Swift-Qwen3.8-27B-Uncensored-BF16
     https://huggingface.co/d0xin/Swift-Qwen3.8-27B-Uncensored-BF16
     A Derivative Work of (2): rank-1 directional residual-stream ablation
     at layer 38 (131 of 1,199 tensors modified, vision tensors unchanged).
     Methodology and validation: ABLITERATION.json,
     STRUCTURAL_VALIDATION.json and INTELLIGENCE_VALIDATION.json in that
     repository.

  4. This work: a Derivative Work of (3). Changes (Apache License 2.0,
     Section 4(b), and Swift Open License v1.0 change notice):
       - model-*.safetensors, model_extra_tensors.safetensors,
         model.safetensors.index.json: the decoder linear layers of (3) are
         quantized to 4-bit integers (W4A16, group size 128, symmetric) with
         Intel AutoRound (128 calibration samples x 2048 tokens). The vision
         tower and the GatedDeltaNet in_proj_a/in_proj_b layers stay BF16.
         embed_tokens is quantized to int8 (group 128, symmetric).
         Swift-Qwen3.8-27B-Uncensored-W4A16: lm_head and the MTP module's
         linear layers are quantized to int8 (group 128, symmetric).
         Swift-Qwen3.8-27B-Uncensored-W4A16-fast: lm_head and the MTP
         module's linear layers are quantized to int4 (group 128,
         symmetric) with GPTQ, calibrated on hidden states captured from
         this model's own generations.
         Both: a 40,960-row draft head (mtp.draft_lm_head.*) is sliced
         from the quantized lm_head for MTP speculative decoding.
       - config.json, quantization_config.json: quantization metadata
         added. generation_config.json: identical to (3).
       - mtp_draft_vocab_ids.pt, draft_vocab_ids.json: added (the token ids
         of the draft head, counted over this model's own outputs).
       - chat_template.jinja: two changes to the Qwen3.8 template.
         (a) Reasoning-effort translation: the OpenAI names "minimal",
         "high" and "max" map to the template's low / xhigh levels, and an
         unknown value no longer raises an error.
         (b) Tool calls whose arguments are a JSON string instead of an
         object render as one <parameter=arguments> block instead of
         failing.
       - README.md: replaced.
       - tokenizer.json, tokenizer_config.json, preprocessor_config.json,
         processor_config.json: re-saved from (3) by transformers 5 (the
         chat template moved to chat_template.jinja); same vocabulary,
         merges and special tokens.
     Tools: the syv-ai/HyperQwen pipeline (run_quant.sh, prepare/,
     drafter/), https://github.com/syv-ai/HyperQwen.

Per Section 4(e) of the Swift Open License v1.0, this distribution includes
a copy of the Base Model License (LICENSE-APACHE-2.0) next to the Swift Open
License v1.0 (LICENSE), because this work incorporates portions of the Base
Model.

Attribution notices from the NOTICE file of ukisai/Swift-Qwen3.8-27b,
reproduced as Section 4(d) of the Swift Open License v1.0 requires:

--------------------------------------------------------------------------
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.
--------------------------------------------------------------------------