Qwen3.8-Flash-Next VRAM requirements at Q4 and 256K context

How this is calculated

The 48 layers repeat one pattern 12 times: 3 Gated DeltaNet layers, then 1 Qwen Sparse Attention layer. That makes 36 fixed state layers and 12 sparse ones, with a hidden size of 2,560. The sparse layers use 2 KV heads at head dimension 256, and each also stores a 128 value index key per token. In FP16 that's 27,648 bytes per token, so 7.2 GB at 262,144 tokens. The DeltaNet layers keep a fixed state of 0.12 GB at batch one, whatever the context. Qwen says the native context can be extended to 1M. At 1M the cache grows to 29.0 GB and the total to 143 GB. At 32K the total is 112 GB. The n-gram table is the unusual part. It's 51B of the 180B, which is 29 GB if it's stored at the same Q4 factor as everything else. Without it, the backbone and speculative layer come to 72 GB of weights. Qwen's notes call the table a candidate for offloading to system memory.

Verdict

Decide about the n-gram table before you pick hardware. If it stays on the GPU, the 256K scenario needs 2 cards of 80 GB by capacity. A single 192 GB card holds the 256K total on paper. A 128 GB unified memory machine holds the 256K total on paper but not the 1M one, and the operating system needs its share too. If the table moves to host memory, the GPU budget falls by roughly the table's share, but the host needs room and you need to measure the cost of the lookups. This estimate leaves out the speculative layer's buffers, prefill workspace, communication between GPUs and any uneven split. Check the engine you plan to use, then test with your real prompts.