Technical Specifications and Architecture
The developers behind GigaChat have announced the release of GigaChat-3.5 Reasoning, a new model in their existing lineup. According to reports on Reddit, the model utilizes a Mixture-of-Experts (MoE) architecture, specifically a 432B-A28B configuration. To address long-context efficiency, the model incorporates Gated DeltaNet technology.
The development process involved training specialized domain experts in fields such as mathematics, coding, and general knowledge using CISPO. These experts were subsequently distilled into a single model through a process of on-policy distillation.
Performance Benchmarks and Efficiency
In preliminary evaluations shared by the developers, GigaChat-3.5 Reasoning reportedly performs at a level close to the DeepSeek V4 Flash Preview. Notably, the developers claim the model achieves these results while utilizing 37% fewer tokens in its reasoning traces compared to its competitor. This efficiency addresses a common criticism of high-reasoning models, which users on Reddit have noted can consume an excessive number of tokens during extended processing tasks.
The model’s weights have been made available on Hugging Face under the MIT license. Users can access the model via the reasoning tab at giga.chat.
Community and Implementation
Following the release, technical discussions on platforms like Reddit and GitHub have focused on local implementation and hardware requirements. A pull request is currently ongoing for llama.cpp to support the new model, according to GitHub documentation.
Local deployment discussions suggest that high-end consumer hardware, such as systems equipped with an NVIDIA RTX 5090 and 64GB of RAM, can run large models like Qwen 3.8 27B at speeds of approximately 70-100 tokens per second when utilizing specific configurations, such as draft models and optimized context settings. However, users have noted that maintaining large context sizes for agentic use cases can be challenging depending on the quantization level used.
While the technical community has expressed interest in the model’s training costs and future iterations, such as a potential 80b MoE version, these remains speculative.