🤖 Implementation Details of GigaChat 3.5 Ultra
Sber has introduced GigaChat 3.5 Ultra — a 432B parameter model trained in FP8 format. The architecture combines MLA with GatedDeltaNet linear layers, which reduces KV cache usage by 4x and increases context by 2.14x. The implementation of MTP heads has increased generation speed by 2.2x.
🌍 The hybrid architecture and the use of FP8 set a new standard of efficiency for ultra-large models, allowing for reduced memory costs and lower inference latency.
👤 Users gain access to a powerful open-source model (MIT license) with high operational speeds that outperforms analogs in mathematics and coding.
Source 1: https://habr.com/ru/companies/sberbank/articles/1055826/ Source 2: https://huggingface.co/collections/ai-sage/gigachat-35