license: other
license_name: qwen-community-1.0
license_link: LICENSE
base_model:

• orcarouter/Qwen3.8-Flash-Next-Uncensored-FP8
• Qwen/Qwen3.8-Flash-Next
  base_model_relation: quantized
  pipeline_tag: image-text-to-text
  library_name: mlx
  language:
• zh
• en
  tags:
• mlx
• omlx
• oq
• mtp
• speculative-decoding
• qwen4_exp
• mixture-of-experts
• abliterated
• uncensored
• traditional-chinese
• zh-tw
• apple-silicon
• 4-bit
  extra_gated_prompt: >-
  This is an abliterated (refusal-removed) model. It will comply with requests the original
  Qwen3.8-Flash-Next refuses. By requesting access you agree to use it only for lawful purposes,
  to add your own safety and moderation layers before any deployment, and to accept full
  responsibility for how you use it and everything it generates.

Qwen3.8-Flash-Next-Uncensored · oQ4e-MTP · zh-TW

A 4-bit MLX build of the abliterated Qwen3.8-Flash-Next for Apple silicon, with the native MTP draft head
kept for speculative decoding, and tuned to answer in Taiwan Traditional Chinese without retraining a single weight.

• Weights: oMLX oQ4e from orcarouter/Qwen3.8-Flash-Next-Uncensored-FP8 — 106.3 GB in 21 shards, native MTP head included (76 tensors).
• zh-TW: 27,640 tokens that contain Simplified-only characters are suppressed at generation time (generation_config.json), and the chat template adds a short Taiwan-usage note.
• Measured on 20 prompts: Simplified characters in Chinese replies went from 2.9 % to 0 %; Mainland terms from 36 to 14; English prompts are still answered in English; speed is unchanged.

The model weights are byte-identical to the plain oQ4e-mtp build; only generation_config.json and chat_template.jinja differ.

Runtime

Requires a runtime with qwen4_exp and its native MTP block — tested with oMLX 0.7.0. In the model settings enable:

On a 128 GB Mac about 76 GB stays resident; the PLE table is read from SSD and cached by the OS, so it is faster when other large models are not loaded at the same time.

What makes it Traditional Chinese

1. Simplified-only tokens are suppressed

generation_config.json carries a suppress_tokens list, which oMLX applies on every request (including the MTP draft/verify path). A token is suppressed when its text contains a character that

• OpenCC s2tw maps to a different Taiwan form, and
• is essentially absent from Taiwanese text: less than 0.5 % as frequent as its Taiwan form in zh-TW Wikinews, or less than 0.3 % in a CC-100 zh-Hant shard (187 M characters).

That selects 1,202 characters → 27,640 tokens (11 % of the vocabulary). Characters that are legitimate in Taiwan usage are kept — 後/后 (皇后), 裡/里 (鄰里), 台, 干, 才, 于 (surname), 云, 准, 托, 余, 范, 征 … — and none of ~70 everyday Traditional characters were caught. The full list is in zhtw/banned_chars.txt; zhtw/freq.py and zhtw/build_ban2.py rebuild it.

Consequence: this build cannot write Simplified Chinese, even when asked. Use the plain oQ4e-mtp build (same weights) if you need that.

2. A Taiwan-usage note in the chat template

The template prepends one English instruction to the system prompt: reply in the user's language, and when replying in Chinese use Taiwan Traditional Chinese and Taiwanese vocabulary (網路, 資料, 使用者, 執行, 快取, 選單, 物件, 軟體, 影片, 預設, 伺服器 …). It is phrased in English on purpose — a Chinese version of the note made the model answer English questions in Chinese.

Turn it off per request with chat_template_kwargs: {"zhtw_style": false}.

Evaluation

20 prompts (zhtw/eval_zhtw.py): 16 Chinese (everyday questions, tech explanations, code with Chinese comments, customer service, poetry), 2 Chinese prompts under an English agent-style system prompt, 2 English prompts. Greedy decoding, max_tokens 400, thinking off, oMLX 0.7.0 on an M5 Max.

*Run on the shared production server while other models were loaded; per-reply speed in the same range (49–68 t/s).

The worst baseline case, 「簡單介紹台灣的健保制度。」, switched to Simplified from the second sentence on (135 of 448 characters: 「台湾政府主导的社会保险制度」); this build answers it entirely in Traditional Chinese. The remaining Mainland terms are vocabulary choices that no character filter can catch (優化 ×2, 運行 ×2, 項目, 程序, 緩存, 菜單); removing them would need fine-tuning on Taiwanese text. Raw outputs: zhtw/eval_baseline.json, zhtw/eval_zhTW.json.

Speed (M5 Max 128 GB, oMLX 0.7.0, MTP on)

The second run is faster because the PLE rows it needs are already in the page cache. With no other large model loaded and the PLE table pre-read into memory, first runs reach the second-run numbers. Prompt processing at 10.5K tokens: ~1,000 t/s.

Conversion

Quantized with oMLX 0.7.0's streaming oQ quantizer from the FP8 checkpoint:

• oQ4e: 4-bit affine, group size 64, with 631 sensitivity-protected tensors raised to 5/6/8 bits; the PLE n-gram tables use group size 32.
• Enhanced mode: importance-matrix calibration streamed one layer at a time (128 samples × 512 tokens, oqe_code_multilingual), see oq_imatrix_report.json.
• preserve_mtp=True: the FP8 checkpoint's own MTP block is kept.

Disclaimer

The upstream model has had its safety alignment removed by abliteration. It will comply with harmful, unethical, or illegal requests that the original model would refuse. It is provided for legitimate 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

Released under the Qwen Community License 1.0 inherited from Qwen/Qwen3.8-Flash-Next; the license text is in LICENSE. Abliteration and the FP8 build are by OrcaRouter. This is a community conversion, not an official Qwen release.

中文摘要

這是 Qwen3.8-Flash-Next 無審查版（abliterated）的 MLX 4-bit 版本，保留 MTP 推測解碼，並針對台灣繁體中文調整，沒有重新訓練任何權重：

• 禁止簡體字：generation_config.json 裡封鎖了含有 1,202 個簡體專用字的 27,640 個 token，回答一律是繁體字。因此這一版無法輸出簡體中文。
• **台灣用語提示：**對話範本會加一段系統提示，要求以中文回答時使用台灣用語（網路、資料、使用者、執行、快取……）。不需要時可傳 chat_template_kwargs: {"zhtw_style": false} 關掉。
• **實測（20 題）：**簡體字比例從 2.9% 降到 0%，大陸用語從 36 次降到 14 次；英文題照常用英文回答，速度不變。
• **需求：**oMLX 0.7.0 以上，開啟 mtp_enabled 與 qwen4_ple_ssd_offload。在 128 GB 的 Mac 上常駐記憶體約 76 GB。