Gigachat 3 5 Reasoning Release
Covered by
reddit.com
Timeline
- GigaChat-3.5 Reasoning released: 432B-A28B MoE under MIT, trained via CISPO distillation
A 432B open-weight reasoning model released under MIT with a claimed 37% token reduction against DeepSeek V4 Flash Preview would shift the cost calculus for anyone serving or studying frontier-scale open models, if the team's own evals hold up.
1 source · thin-sourcing, unconfirmed
All sources (1)
- r/LocalLLaMA2026-09-10