Deepseek V4.1 Flash is 748B total/552B base

Inspection of the safetensors shows Deepseek V4.1 Flash’s main model is about 551.566B parameters (40 layers), with FFN experts totaling 543.582B and other parts 7.984B; additionally an engram of ~196.929B, DSpark/MTP ~14.225B and a vision encoder ~0.485B, summing to roughly 748B. Other reported sizes (e.g., 485B, 522B) stem from differing counting methods or misreads.

• NVIDIA advances native GPU programming in Rust2026-09-17
• Offload most of Qwen3.8 KV cache to RAM2026-09-16
• Censorship Begins on HuggingFace2026-09-15
• Aurora1.0-150M Released2026-09-14
• DeepSeek V4.1 Flash tops AA's new benchmark2026-09-14