trjxter/Qwenseek-3.8-27B-CyberLite-BF16 - Каталог нейросетей
Генерация текста

trjxter/Qwenseek-3.8-27B-CyberLite-BF16

Добавлено:
trjxter/Qwenseek-3.8-27B-CyberLite-BF16

The original BF16 release had extra wrapper prefixes in its Safetensors headers and index. Stock loaders could therefore leave language and vision layers unloaded. This release repairs 1,183 names across 11 shards using these exact rules: All 1,199 tensor names and shapes now match both Qwen/Qwen3.8-27B and the unsloth/Qwen3.8-27B base at the revisions recorded in the validation report. The tensor payloads, dtypes, shapes, and offsets are unchanged; payload preservation was checked by SHA-256 after rereading every rewritten shard. lmhead.weight` and the 15 MTP tensors are unchanged. Existing FP32 tensors are retained as saved. The index parameter count also now includes the previously restored MTP tensors (27,781,427,952 parameters in total); its byte count was already correct. Existing users: refresh the checkpoint from the current main revision and restart the loader. Download the updated shards together with the updated index; replacing only the index does not repair an older local checkpoint. All 18 published GGUF variants passed a separate structural audit. The pinned converter already removes repeated languagemodel.` prefixes, so the GGUF files do not require a rename or…

Модальности:
Генерация текста

Области применения:
Генерация кода Логика и рассуждение Диалог / чат


Задача: Генерация текста
Автор: trjxter
Теги: qwen3_5, image-text-to-text, qwen3.8, unsloth, lora, qlora, sft, bf16
Лайков: 3  |  Загрузок: 140

Открыть на HuggingFace →

Описание основано на материалах HuggingFace. Перевод выполнен автоматически.