摘要

Qwen3.8-Flash-Next 是一个实验性开源权重多模态模型，预演 Qwen4 计划采用的架构：125B 语言模型（每 token 激活 6B），外加 51B n-gram 嵌入表与 4B 多 token 预测头（BF16 合计约 180B）。注意力为混合式——Gated DeltaNet 线性注意力为主、每 4 层插一层 Qwen Sparse Attention；原生 262,144 token 上下文，YaRN 可扩到 1M。模型卡 benchmark 上领先 Qwen3.8-27B 与 DeepSeek-V4-Flash-0731（DeepSWE 1.1 58.7、SWE-bench Pro 62.5、LiveCodeBench v6 91.9、GPQA Diamond 91.7）。以 qwen-community-1.0 自定义许可发布，提供 Transformers / vLLM / SGLang 三种后训练权重形态；官方定位中 Qwen3.8-Flash 是它的生产对应版本。

为什么重要
对服务团队来说，这是一周内第二个把筹码押在「混合线性注意力降长上下文成本」上的前沿实验室开源权重——GLM-5.3-Flash 在 8/25 落地了同一思路。把它当架构信号而不是生产可用版本：官方明确标注实验性质、许可证是自定义而非 MIT/Apache，且默认 1M 上下文的生产版 Qwen3.8-Flash 才是受支持的对位产品。
技术细节
架构 125B 语言模型（激活 6B）+ 51B n-gram 嵌入（bigram/trigram 查表）+ 4B 多 token 预测；BF16 合计约 180B；48 层循环模式：3 x（Gated DeltaNet -> MoE）+ 1 x（Qwen Sparse Attention -> MoE）
注意力 混合式：Gated DeltaNet 线性注意力（48 V / 16 QK 头，dim 128）+ 每 4 层一次 Qwen Sparse Attention（24 Q / 2 KV 头，dim 256；索引预算 512 blocks / 2048 tokens）
MoE 512 专家，每 token 激活 10 路由 + 1 共享；Gated Residual 4 分支，瓶颈秩 320
上下文 原生 262,144；YaRN 可扩至 1M
模态 图-文输入的多模态；默认思考模式（effort xhigh/medium/low）
训练 Muon 与 AdamW 按权重类别分工；不做 batch-size warmup；更大学习率
Benchmark DeepSWE 1.1 58.7、SWE-bench Pro 62.5、SWE-bench Multilingual 81.0、NL2Repo-Bench 48.1、Agents' Last Exam 51.2、GPQA Diamond 91.7、LiveCodeBench v6 91.9；视觉：AndroidWorld 84.5、LVBench 76.6——对比 Qwen3.8-27B、Qwen3.7-Plus、DeepSeek-V4-Flash-0731、Claude-Opus-4.6
许可证 自定义 qwen-community-1.0（license: other）——商用前先核对条款
采用 截至 2026-08-30T00:00Z：主仓 52,341 下载（30d）+ 4,287 likes；FP8 变体 44,281 下载；生产对位版本 Qwen3.8-Flash 自发布周起在 QwenCloud / Model Studio / OpenRouter 提供（$0.15 / $0.47 每百万 token）
后续更新
2026-08-30 Production counterpart verified (resolves the prior watch item): Qwen3.8-Flash is the API-only production version of Flash-Next — live on QwenCloud since launch week with a 1M-token default context and official built-in tools; served via Alibaba Cloud Model Studio (USD 0.15/M input <=1M tokens, 0.47/M output, context-caching discount) and OpenRouter (0.15/0.47). No separately licensed open weights beyond Flash-Next itself (custom qwen-community-1.0). Adoption re-check 2026-08-30: main repo 52,341 downloads + 4,287 likes (was 4,213 on 8/29; the 30d download pool is unchanged at 52,341 — treat as a static window, not zero growth). Recommendation stays WATCH (experimental preview; production path is the API).
2026-08-31 Adoption check @2026-08-31T00:00Z: 121,976 downloads (+133% vs the 52,341 @8/30 sample, ~7 days after listing) + 4,384 likes (+97) — the display counter finally moved after appearing frozen between the 8/29 and 8/30 samples. Sustained-velocity data point for trend #1 criterion (c). Recommendation stays WATCH.
2026-09-02 adoption check @2026-09-02T00:0xZ: 207,941 downloads (+70% vs 121,976 @8/31, ~9 days after listing) + 4,635 likes; FP8 variant 130,451; unsloth GGUF derivative 431,339 (repo created 8/26). slotstream (Show HN 9/1, 147 pts) runs the 104GB 4-bit weights on a 48GB M5 Pro at ~12 tok/s warm decode (32GB peak, SSD streaming, single Swift binary, Ollama/OpenAI-compatible APIs) — independent community serving tooling, though still not a benchmark reproduction. Architecture now documented in the Qwen3.8-Next design paper (arXiv 2608.30320, ev-20260901-03): 125B/6B active + 51B n-gram embeddings, ~1/9 training FLOPs vs the 397B-A17B predecessor at comparable quality. Stays WATCH
2026-09-05 Adoption check @2026-09-05: 351,374 downloads (+69% vs 9/2) + FP8 repo 186,676 (+43%) + 4,871 likes; unsloth/Qwen3.8-Flash-Next-GGUF at 702,251 — now exceeds Qwen's own main repo; nvidia/Qwen3.8-Flash-Next-NVFP4 created 9/2 (NVIDIA publishing an NVFP4 derivative); mudler APEX GGUF derivative 9/4. Quantization recipe research catching up: Minima NVFP4 W4A4 across a Qwen3.8-27B hybrid incl. GDN layers (ev-20260905-03). Stays WATCH
2026-09-14 2026-09-14 report run: SGLang main merged a reproducible single-DGX-Spark path for NVIDIA's NVFP4 derivative (#39126). The narrow 200-question GSM8K check reported 97.5%/97.0% with/without MTP and 92.87/89.53 tok/s, but this is TP=1, main-branch-only and not a controlled comparison. Recommendation stays WATCH.
标签
open-weightsarchitecture-previewlinear-attentionmoelong-contextmultimodal