Qwen3.8-Flash-NextOpen Weight

Qwen3.8-Flash-Next, released August 26, 2026, was Alibaba's public preview of the architecture it intended to build Qwen4 on, put out as downloadable weights under a custom qwen-community licence. It was a natively multimodal mixture-of-experts model with 125 billion parameters and only 6 billion active per token, carrying a further 51 billion parameters of N-gram embeddings — a way of adding capacity that costs less computation than adding experts and is easier to offload from accelerator memory. Its 262K-token native context window extended to roughly 1M through YaRN.

The rest of the design was equally experimental. Gated DeltaNet was paired with Qwen Sparse Attention, which selects context at the level of micro-blocks rather than individual tokens to cut latency on the long agentic prompts Alibaba said were coming to dominate real usage; a gated residual stream added separate read and write gates; and the training recipe split the Muon and AdamW optimisers across different weight categories and started at the target batch size instead of warming up to it. Alibaba put the training cost at roughly one-ninth that of Qwen3.7-Plus, the 397B proprietary model it was benchmarked against.

The launch table was built to show a 6B-active model beating much larger ones. It scored 62.5% on SWE-bench Pro and 81.0% on SWE-bench Multilingual, 58.7% on DeepSWE 1.1, 73.9% on CoWorkBench and 55.7% on JobBench — the last of those 19 points above the 36.6% Alibaba reported for Claude Opus 4.6 Max and 28 above Qwen3.7-Plus. It posted 91.7% on GPQA Diamond and 91.9% on LiveCodeBench v6, and 35.9% on HLE, the one language row where Opus 4.6 Max led at 40.0%. Computer use was the visible gap: 19.4% on the binary scoring of OSWorld 2.0, level with the smaller Qwen3.8-27B. Alongside the weights Alibaba announced a hosted production model, Qwen3.8-Flash, at $0.16 per million input tokens and $0.47 per million output on its QwenCloud API.