> New: Velocity with the Sharp chat template > at xhigh keeps its accuracy and thinks 14–26% less on HMMT, C-Eval and IFBench. > Results on the LoRA card. GGUF builds of Qwen3.8-27B-Velocity (Qwen3.8-27B with the Velocity LoRA merged), made with a code-focused importance matrix and a measured per-tensor bit allocation. V5 and V6 are our measured mixes; V5 keeps clearly more of the model than Q5KM at the same size; V6 edges out Q6_K. Measured on the full-precision model (BF16), all at xhigh, on the same questions and token budgets. Each cell is accuracy / thinking tokens per answer / thinking vs stock. Swift 1.5 (UkisAI) is another Qwen3.8-27B tune that cuts thinking. — Token budget per answer (UkisAI’s published configs, the same for every model): AIME and HMMT 250,000, IFBench 81,920, C-Eval 16,384, LiveCodeBench v6 32,768. — Budget hits: none on AIME, HMMT or IFBench for Velocity; C-Eval 4 of 1,346 (stock 2.0%). On LiveCodeBench Velocity ran out on 61 of 350 answers, Swift 1.5 on 83 and stock on 118, so much of the LiveCodeBench gap is about finishing inside 32k. — Stock is each question’s average over UkisAI’s 5 published stock runs (our own 2 on LiveCodeBench v6). — Swift 1.5…
Модальности:
Генерация текста
Области применения:
Логика и рассуждение Диалог / чат
Задача: Генерация текста
Автор: joycx
Теги: gguf, llama.cpp, imatrix, qwen3_5, reasoning, endpoints_compatible, conversational
Лайков: 3 | Загрузок: 3,051
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
