license: other
license_name: swift-open-license-1.0
license_link: LICENSE
base_model:

• jessedye90/qwen3.8-flash-next-swift-uncensored
  base_model_relation: quantized
  pipeline_tag: image-text-to-text
  library_name: mlx
  language:
• zh
• en
  tags:
• mlx
• omlx
• mtp
• speculative-decoding
• qwen4_exp
• qwen3_8
• mixture-of-experts
• abliterated
• uncensored
• traditional-chinese
• zh-tw
• apple-silicon
• 4-bit
  extra_gated_prompt: >-
  This model has had its safety refusals removed (abliteration). It will comply with harmful requests that the
  original model refuses. By requesting access you agree to use it only for lawful purposes, to add your own safety
  and moderation layers before any deployment, to accept full responsibility for how you use it and everything it
  generates, and to comply with the Swift Open License v1.0 (including its commercial-use limitation) and the Qwen
  Community License 1.0.

Swift 1.5 Qwen3.8-Flash-Next Uncensored · MLX 4-bit MTP · zh-TW

An MLX build of jessedye90/qwen3.8-flash-next-swift-uncensored
(UkisAI's Swift 1.5 fine-tune of Qwen3.8-Flash-Next, with the refusal direction removed) for Apple silicon. The native
MTP head is kept for speculative decoding, the n-gram (PLE) table is included so the checkpoint is self-contained, and
generation is tuned to Taiwan Traditional Chinese without retraining any weight.

• Lossless where it matters. The source is GPTQ W4A16. Its INT4 routed experts, attention and DeltaNet projections,
  the INT8 lm_head and the INT4 MTP drafter modules are repacked, not re-quantized: the integer codes are copied
  bit-for-bit into MLX affine quantization with the same bits and group size, and only the fp16 scales are rounded to
  bfloat16. The source's abliteration was made inside those INT4 grids; re-quantizing would partly undo it, repacking does not.
• zh-TW: 27,640 tokens containing Simplified-only characters are suppressed (generation_config.json), and the chat
  template adds a short Taiwan-usage note.
• Measured on 20 prompts: Simplified characters 3.0 % → 0 %, Mainland terms 17 → 7, English prompts still answered in English.
• Size: 102 GB in 22 files (70 GB weights + 32 GB PLE table). On a 128 GB Mac oMLX keeps ~66 GB resident and reads the PLE table from SSD.

Runtime

Requires a runtime with qwen4_exp and its native MTP block — tested with oMLX 0.7.0 on an M5 Max (128 GB).
In the model settings enable:

Conversion

conversion/convert_swift.py streams the source repo one shard at a time and writes the oMLX qwen4_exp layout
(per-expert tensors stacked into switch_mlp). Per tensor family:

GPTQ v1 zero points (stored as 2^(bits-1) − 1) become MLX biases −scale × 2^(bits-1); scales are rounded from float16 to bfloat16 (relative error ≤ 0.2 %).
conversion/swift_conversion_report.json lists every module with its source format, bits, group size, and cosine
similarity to the corresponding tensor of our oQ4e build of OrcaRouter's abliterated Qwen3.8-Flash-Next (0.96–1.0;
routed experts ≈ 0.99, the expected gap between two 4-bit quantizations of closely related weights).

The source repo does not include the n-gram (PLE) table; the copy here (MLX 4-bit, group 32) comes from our oQ4e build of
orcarouter/Qwen3.8-Flash-Next-Uncensored-FP8.
Tokenizer files are the original Qwen3.8-Flash-Next ones (same 248,044-token vocabulary as the source; the source's
re-serialized tokenizer.json drops \p{M} from the pre-tokenizer regex).

What makes it Traditional Chinese

Same method as KaedeTai/Qwen3.8-Flash-Next-Uncensored-oQ4e-mtp-zhTW:

1. Simplified-only tokens are suppressed. A token is in suppress_tokens when it contains a character that OpenCC
   s2tw maps to a different Taiwan form and that is essentially absent from Taiwanese text (Wikinews zh-TW and a
   CC-100 zh-Hant shard). 1,202 characters → 27,640 tokens. Characters legitimate in Taiwan usage (后, 里, 台, 干, 才,
   于, 云, 准, 托, 余, 范, 征 …) are kept. List and scripts: zhtw/. Consequence: this build cannot write Simplified Chinese.
2. A Taiwan-usage note in the chat template, phrased in English so English questions are still answered in English.
   Turn it off per request with chat_template_kwargs: {"zhtw_style": false}.

Evaluation

20 prompts (zhtw/eval_zhtw.py): 16 Chinese, 2 Chinese under an English agent-style system prompt, 2 English.
Greedy decoding, max_tokens 400, thinking off, oMLX 0.7.0, M5 Max, MTP on, on a shared server with another model loaded.

Raw outputs: zhtw/eval_baseline.json, zhtw/eval_zhTW.json. Refusal behaviour was not re-measured after conversion;
see the source model card for its AdvBench / XSTest results.

Disclaimer

Safety alignment has been removed by abliteration. The model will comply with harmful, unethical or illegal requests
that the original model refuses. It is provided for research, red-teaming and private use; you are responsible for how
you use it and for everything it generates. Add your own safety layers before exposing it to anyone else.

License and attribution

• Swift Contribution by UkisAI: Swift Open License v1.0 (LICENSE), including its commercial-use limitation —
  commercial use by an entity with US$1,000,000 or more in annual gross revenue needs a separate Swift Enterprise License
  from UkisAI.
• Base model Qwen3.8-Flash-Next, Copyright (c) 2026 Qwen: Qwen Community License 1.0 (LICENSE-QWEN).
• Abliteration by jessedye90 using OrcaRouter's refusal direction. Attribution and change notices, including the ones
  for this conversion, are in NOTICE.

This is a community conversion. It is not made or endorsed by UkisAI, Qwen, OrcaRouter or jessedye90.

中文摘要

這是 UkisAI Swift 1.5（Qwen3.8-Flash-Next 微調版）無審查版的 MLX 4-bit 版本，保留 MTP 推測解碼，並針對台灣繁體中文調整：

• **不重新量化：**原始檔的 GPTQ INT4／INT8 權重（routed experts、注意力、DeltaNet、lm_head、MTP）整數碼逐位元搬進 MLX 格式，只有 scale 從 fp16 轉成 bf16，保留原作者在 INT4 格點上做的去審查修改。
• **自帶 PLE 表：**原 repo 沒附 n-gram 表，這裡一併提供，下載即可用。
• **禁止簡體字＋台灣用語提示：**封鎖 27,640 個含簡體專用字的 token，對話範本加上台灣用語提示。實測簡體字 3.0% → 0%，大陸用語 17 → 7，英文題照常用英文回答。這一版無法輸出簡體中文。
• **需求：**oMLX 0.7.0 以上，開啟 mtp_enabled 與 qwen4_ple_ssd_offload。128 GB 的 Mac 上常駐記憶體約 66 GB。
• **授權：**Swift Open License v1.0（年營收 100 萬美元以上的公司商用需另取授權）＋ Qwen Community License 1.0。