Qwen3.8-Flash-Next vs Qwen 3.5 122B-A10B (MoE)

Analysis

Qwen 3.5 122B-A10B and Qwen3.8-Flash-Next are six months apart from the same team, and underneath the size difference they share an unusual amount of skeleton: both are 48-layer hybrid MoE stacks built from twelve repeats of three Gated DeltaNet layers followed by one full-attention-family layer, and both cache with the identical 2 KV heads by 256 head dim. The only thing that changed in that one-in-four slot is what kind of layer sits there: Qwen 3.5 122B-A10B uses plain Gated Attention, Qwen3.8-Flash-Next replaces it with the newer Qwen Sparse Attention (QSA).

Because the caching layer count, ratio, and width are identical, the two models' KV cache scales as exactly the same function of context length, byte for byte, despite one being a six-months-newer preview architecture and the other a shipped production release. What differs is everything upstream of that cache: Qwen3.8-Flash-Next routes through 512 experts (10 routed plus 1 shared) against Qwen 3.5 122B-A10B's 256 (8 routed plus 1 shared), and activates fewer parameters per token doing it, 6B against 10B, while carrying nearly 50% more total parameters, 180B against 122B. Most of that extra mass isn't backbone at all: Qwen3.8-Flash-Next's real 180B total includes a 51B-parameter n-gram embedding table its own technical report says is meant to run from host RAM, not VRAM, a wrinkle Qwen 3.5 122B-A10B's conventional weight file doesn't have. On the two benchmark fields both report, the newer model leads clearly despite activating little more than half the parameters: GPQA Diamond 91.7 against 86.6, and LiveCodeBench v6 91.9 against 78.9. Qwen 3.5 122B-A10B answers back with everything Qwen3.8-Flash-Next doesn't have yet: real community GGUF quants, an Ollama tag, and the simpler Apache 2.0 license, against Qwen3.8-Flash-Next's Community 1.0 terms, which require a separate license for any Model-as-a-Service or 'AI Work Assistant' business built on it. Both accept text, image, and video input.

Bottom line: Qwen 3.5 122B-A10B is the model to actually run today: it fits an 80GB card at its recommended Q3_K_M, has a real quant ladder and Ollama support, and scores respectably on the benchmarks both share. Qwen3.8-Flash-Next scores higher on those same benchmarks at less than half the active compute, but it is explicitly a preview of Qwen4's likely architecture, needs a 141GB-or-larger card at its own recommended Q4_K_M (vLLM and SGLang can offload the 51B n-gram table to host RAM instead, which lowers that bar, but neither a GGUF build nor a confirmed community quant exists yet). Pick Qwen3.8-Flash-Next to see where Qwen's hybrid-MoE recipe is headed next; pick Qwen 3.5 122B-A10B if you actually need to load something this week.