GigaChat 3.5 arrives as a 432B mixture-of-experts model

The team behind GigaChat has published GigaChat3.5-432B-A28B on Hugging Face, a large mixture-of-experts language model aimed at instruction-following across multiple languages. According to the model page, it carries 432 billion total parameters but activates only about 28 billion for any given token.

That sparse design is the whole point. Instead of running every parameter for every request, an MoE model routes each token through a small subset of specialized "experts," which keeps inference costs closer to a mid-sized dense model while retaining the capacity of a much larger one. GigaChat 3.5 pairs this with hybrid attention, a technique meant to trim the memory and compute overhead that normally grows with longer inputs.

Why it matters

Open-weight MoE releases at this scale remain relatively rare, and each one gives practitioners another option outside the handful of frontier labs. The key details here:

• 432B total parameters, with roughly 28B active per token
• Multilingual instruction tuning rather than an English-only focus
• Hybrid attention to manage the efficiency trade-offs of a very large model

A few specifics are still unstated on the release page, notably the context window and the exact licensing terms, both of which will shape how usable the model is in production. For now, GigaChat 3.5 lands as a notable addition to the growing roster of large sparse models available to download and run.