LaunchAI ModelsSeptember 10, 2026

GigaChat-3.5 Reasoning released as 432B-A28B MoE

GigaChat-3.5 Reasoning is a 432B-parameter MoE with 28B active parameters, using Gated DeltaNet for long-context efficiency. Domain experts (code, math, general) were trained with CISPO, then distilled into one model via on-policy distillation.

1 source

More stories today

Open the live feed
GigaChat-3.5 Reasoning released as 432B-A28B MoE — AIBriefs