Qwen3.8-Flash-Next: Open-Weight Preview of Qwen4's Hybrid Attention on GB300 NVL72

Alibaba released Qwen3.8-Flash-Next as an open-weights early preview of the Qwen4 architecture. NVIDIA positions the release for context-intensive agentic coding and documents Day-0 support in SGLang, vLLM, and TensorRT LLM.¹²

Context

The model is a multimodal mixture-of-experts with 176B total parameters, 6B active per token, and a 262,144-token native context extensible to 1M tokens with YaRN. Its architecture interleaves three Gated DeltaNet layers with one Qwen Sparse Attention layer, compressing history into a fixed-size recurrent state while retaining block-level retrieval over the full context.¹

Assessment

The central constraint shift is the removal of KV-cache growth: GDN layers continuously compress historical context into a fixed recurrent state, and QSA selects relevant micro-blocks rather than paying attention over the entire sequence. Alibaba-published benchmarks report up to 7.6x prefill and 4.9x decoding kernel speedups over full attention, and an 8.6x prefill throughput advantage over Qwen3.7-Plus in a 1M-token, 90% prefix-cache-hit serving test.¹

Caveats

These are vendor-reported numbers, not independent reproductions; the throughput figures refer to a reference GB300 NVL72 rack and may not generalize to smaller deployments. A hands-on report describes the model as '125B tokens' with 6B active, a discrepancy from the stated 176B total that must be resolved before size-efficiency comparisons. Independent reproductions of the 1M-token benchmarks and third-party latency measurements would materially change the assessment.¹²