Sber has introduced GigaChat 3.5 Ultra — a massive model with 432B parameters optimized for high efficiency and speed. By combining an innovative architecture with FP8 training methods, the model demonstrates results on par with DeepSeek V3.2, surpassing it in coding and mathematics tasks.
What Happened
Sber introduced GigaChat 3.5 Ultra, a 432B parameter model trained in FP8 format. The architecture combines Multi-Head Latent Attention (MLA) with GatedDeltaNet linear layers, which reduces KV cache usage by 4x and increases the context window by 2.14x. The use of Multi-Token Prediction (MTP) heads for speculative decoding increased generation speed by 2.2x, while throughput under load increased by 20%. The model supports over 600 languages and is released under the MIT open license.
Context
The development aims to solve latency issues during the inference of ultra-large models and to improve their economic efficiency. The hybrid architectural approach and the transition to low-precision FP8 training allow for high performance with significantly lower memory and computational resource costs compared to traditional methods.
Why It Matters for the Industry
The transition to a hybrid architecture (MLA + GatedDeltaNet) and the use of FP8 training sets a new efficiency standard for ultra-large models, making high-performance LLMs more accessible for production. The application of MTP heads directly addresses inference latency issues, and the MIT license release lowers barriers to creating complex AI agents and high-load systems.
Why It Matters for Users
Developers and engineers gain access to a powerful open-source model that excels at complex coding and mathematical tasks while operating faster and more efficiently than heavier counterparts. This allows for the deployment of advanced AI tools locally or in private clouds, reducing total cost of ownership and latency.
Sources
Author
Look at AI, Editorial Team