This repository contains a highly optimized TQ1 quantized version of the official tiiuae/Falcon3-10B-Instruct model. It is formatted explicitly for the ATLAS Engine ecosystem, designed for native, ultra-low-latency CPU inference without any GPU requirement. > Packed using the unified packtoatlas.py toolchain (v2.10.0) with BF16 weight scale correction. During pre-release evaluation (v2.10.0), this quantized derivative demonstrated correct convergence: — T=0 (argmax): «The capital of France is Paris.» — correct deterministic output — T=0.7 (sampling): Coherent structured generation with sensible continuation > Note on scale mathematics: the legacy dequantization path divides by the scale factor rather than multiplying. Since this is a constant across all logits for any given output row, the relative probability distribution remains identical under softmax normalization — no effect on output quality. — Engine source code: github.com/xxxn3m3s1sxxx/ATLAS-TQ10 — Original model: tiiuae/Falcon3-10B-Instruct — TII Falcon3 License**: tiiuae/falcon3-license This is a quantized derivative work based on Falcon3 architectures developed by the Technology Innovation Institute (TII). By…
Модальности:
Генерация текста
Области применения:
Следование инструкциям
Задача: Генерация текста
Автор: xxxn3m3s1sxxx
Теги: atlas, falcon3, ternary, quantized, tq1, cpu-optimized, cpu-inference, bitnet
Лайков: 3 | Загрузок: 26
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
