voves/GigaChat3.5-432B-A28B-NVFP4

I took GigaChat 3.5 Ultra (ai-sage/GigaChat3.5-432B-A28B) — a 432B Mixture-of-Experts model with 28B active parameters, built on a custom hybrid architecture pairing Multi-head Latent Attention (MLA) with GatedDeltaNet linear-attention layers, GatedNorm, and two Multi-Token Prediction heads and compressed it from its native BF16 release down to NVFP4. Two things made this non-trivial. First, the MoE routing: with 256 routed experts and top-8 selection, an FP4 quantization that corrupts the router, shared expert, or the MLA projections collapses quality silently — so I excluded lm_head, embeddings, the shared expert, all gate/router modules, the low-rank q/kv projections, and the first 3 dense + last 2 layers from quantization. Second, the calibration data: an English corpus would have left Russian token activations and expert routing patterns under-sampled, so I built a 2k-sample Russian mix dataset tokenized at 4096 ctx to cover the model's YaRN-scaled long context.