Qwen3.8-Flash-Next

Alibaba's August 26, 2026 open-weight architecture preview — explicitly an early look at the stack Qwen4 will be built on, released early so the community can examine it, exactly as Qwen3-Next was. The headline is efficiency: 6B activated parameters, 8.6x the prefill throughput of Qwen3.7-Plus at 1M context under a 90% prefix-cache hit rate, and $0.15/$0.47 per Mtok for the production `qwen3.8-flash` SKU on QwenCloud (which defaults to 1M context and built-in tools). The 51B N-gram embedding table is the notable trick — capacity added at almost no per-token compute, stored in host memory and prefetched asynchronously. Results are strong for the activated size (SWE-bench Pro 62.5 beats Claude Opus 4.6 Max's 53.4; SWE-bench Multilingual 81.0) but the ceiling shows on long-horizon work: DeepSWE 1.1 58.7 and OSWorld 2.0 19.4 are far off the frontier. Rows are taken from the post-trained model card at the native 262K context, not the 1M YaRN extension; the in-house CoWorkBench, RecreationBench and ClawEval-MM benchmarks are deliberately not filed as catalog benchmarks.