EXL3 quantization of orcarouter/Qwen3.8-Flash-Next-Uncensored (BF16, tensor-level abliteration). 125B-A6B Qwen4-preview MoE at 2.05 bits per weight. All 512 experts fit on 2×24 GB GPUs with no CPU offload. > Research-only release. Safety alignment has been largely removed > (see Disclaimer). You carry full responsibility for use. See License. Only plain quants at 2.05 bits per weight are available in this repository. No other bitrates or quant types are published here. Release status: this is the only planned release. No further quants will be made from this source in this repository. Stable over 5 back to back runs (range 1.3 tok/s). TTFT ~0.2 s warm. Envelope: -gs 20,23.5 -cs 98304 -cq 8, thinking ON. Drafting note: MTP speculative drafting measured 52 to 55 tok/s on these weights, against ~72 without. Draft acceptance on the abliterated weights costs more than it saves, so this release is served with -dm none. Outputs match either way (drafting changes speed only, never content). 40-prompt scripted battery (temperature: 0), all verdicts automated. 40/40 PASS, rechecked under the shipping -dm none config: — Rare words, idioms, multilingual (ES/FR/DE), low-resource (Swahili,…
Модальности:
Генерация текста
Области применения:
Логика и рассуждение Диалог / чат Вызов функций (Tool use)
Задача: Генерация текста
Автор: VinceTrune
Теги: qwen4_exp, exl3, abliterated, uncensored, moe, flash-next, reasoning, tool-use
Лайков: 3 | Загрузок: 21
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.