[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"news-slug-qwen3-8-27b-open-weights-release":3,"news-related-cb64371f-62b6-473d-8150-b576001d3f56":38},{"id":4,"title":5,"summary":6,"content":7,"original_url":8,"source_id":9,"tags":10,"translations":24,"news_slug":31,"published_at":32,"created_at":33,"modified_at":34,"is_published":35,"publish_type":36,"image_url":14,"view_count":37},"cb64371f-62b6-473d-8150-b576001d3f56","Qwen3.8-27B 开源权重上线:单卡跑得动的 Qwen3.8,还塞了个视觉编码器","阿里 Qwen3.8-27B 权重已于 8 月 13-14 日落地 Hugging Face,采用 Apache 2.0 许可。28B dense 架构、原生多模态(附带视觉编码器)、262k 上下文可经 YaRN 扩展至 1M。SWE-Bench Pro 61.7%、LiveCodeBench 90.3%,4-bit 量化后约 14-17GB 显存即可本地运行,24GB 消费级显卡就能跑——与 2.4T 参数、API-only 的 Qwen3.8-Max 构成开源\u002F托管双线布局。","8 月 3 日 Qwen3.8-Max 发布时,阿里承诺约一周内放出配套开源权重。8 月 13-14 日,承诺兑现:[Qwen3.8-27B](https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen3.8-27B) 权重正式登陆 Hugging Face,Apache 2.0 许可,距旗舰发布 11 天。\n\n## 规格超预期\n\n这是一个 28B 参数的 dense 模型,采用混合注意力架构。规格表里有两个此前没人承诺过的点:**原生多模态**——自带视觉编码器;**262k 原生上下文**,通过 YaRN 可扩展到 1M token。前者意味着开源社区拿到的不只是文本模型,而是一个可直接做图文任务的多模态基座。\n\n## 跑分:小模型的「越级打」\n\n按 [AI Release Tracker](https:\u002F\u002Faireleasetracker.com\u002Fmodel\u002Fqwen\u002Fqwen3.8-27b) 整理的发布数据:编码向 SWE-Bench Pro 61.7%(该榜第一 Claude Fable 5 为 80.3%)、DeepSWE 1.1 42.2%、LiveCodeBench 90.3%(第一为 DeepSeek-V4-Pro 93.5%)、Terminal-Bench 2.1 73%;办公向 CoWorkBench 70.7%——阿里报告里 Opus 4.6 Max 是 68.2%;GPQA Diamond 89.2%。在该站追踪的全部模型中,Qwen3.8-27B 在 NL2Repo-Bench、QwenSWEBench、CoWorkBench、IFBench、Agent's Last Exam(score)五个榜上暂列第一。当然,这些是官方发布数字,独立复现还要等。\n\n## 显存账单\n\n[Yotta Labs](https:\u002F\u002Fwww.yottalabs.ai\u002Fpost\u002Fqwen-3-8-27b-specs-hardware-requirements-how-to-run-2026) 的测算:BF16 约 56GB(H100\u002FH200 级),FP8 约 28GB(48GB 卡),4-bit 量化约 14-17GB——Unsloth 给出的本地量化方案即约 17GB,意味着一张 24GB 的 RTX 4090 就能跑。对比同为国产旗舰的 Kimi K3:2.8T 参数、自托管起步 1.56TB 权重内存、集群规模。这就是 27B 存在的意义——不是比谁大,是比谁能落地。\n\n## 所以呢\n\nQwen3.8-Max 走 API(每百万 token 2\u002F6 美元),27B 走开源,阿里把同代能力切成了两条分发线。前代 Qwen3.6-27B 本就是社区本地 coding\u002Fagent 最受欢迎的模型之一,这代加上多模态和 262k 上下文,单卡自托管的门槛没变、能力上限抬了一截。对自建推理的团队,这是本月最值得跑一遍 eval 的新权重;对观望的人,答案也很简单:下载是 Apache 2.0,跑不跑得动你的活,试一下就知道。","https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen3.8-27B","c36a21ac-2a77-421b-9519-1e150695732a",[11,15,18,21],{"id":12,"name":13,"slug":13,"description":14,"color":14},"7e89b5cc-57db-4f37-bc6d-28919a73931c","model-release",null,{"id":16,"name":17,"slug":17,"description":14,"color":14},"499f4b56-819d-49a3-9609-33e775143b86","multimodal",{"id":19,"name":20,"slug":20,"description":14,"color":14},"b9bd9039-fcdb-41a8-b85b-fc1587def2b9","open-source",{"id":22,"name":23,"slug":23,"description":14,"color":14},"c187600e-804c-4697-b828-1e4330e0eb10","qwen",[25],{"id":26,"lang":27,"title":28,"summary":29,"content":30},"245ed14e-ed8d-45a9-9d9e-51534f340697","en","Qwen3.8-27B open weights: single-GPU, with vision encoder","Alibaba's Qwen3.8-27B weights landed on Hugging Face on August 13-14 under Apache 2.0. It's a 28B dense, natively multimodal model (vision encoder included) with 262k native context extensible to 1M via YaRN. Scoring 61.7% on SWE-Bench Pro and 90.3% on LiveCodeBench, it runs locally in roughly 14-17GB of VRAM at 4-bit — a single 24GB consumer GPU suffices — forming a two-track open\u002Fhosted split alongside the 2.4T, API-only Qwen3.8-Max.","When Qwen3.8-Max launched on August 3, Alibaba promised to ship companion open weights within about a week. On August 13-14, that promise was kept: [Qwen3.8-27B](https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen3.8-27B) weights officially landed on Hugging Face under Apache 2.0, eleven days after the flagship.\n\n## Specs That Over-Delivered\n\nThis is a 28B-parameter dense model with hybrid attention. Two things in the spec sheet nobody had promised: **native multimodality** — a vision encoder is included — and **262k native context**, extensible to 1M tokens via YaRN. The former means the open community isn't just getting a text model, but a multimodal base ready for image-text tasks.\n\n## Benchmarks: A Small Model Punching Up\n\nPer data compiled by [AI Release Tracker](https:\u002F\u002Faireleasetracker.com\u002Fmodel\u002Fqwen\u002Fqwen3.8-27b) at release: coding-wise, SWE-Bench Pro 61.7% (leader Claude Fable 5 sits at 80.3%), DeepSWE 1.1 42.2%, LiveCodeBench 90.3% (led by DeepSeek-V4-Pro at 93.5%), Terminal-Bench 2.1 73%; office-wise, CoWorkBench 70.7% — versus the 68.2% Alibaba reported for Opus 4.6 Max; GPQA Diamond 89.2%. Among all models tracked on that site, Qwen3.8-27B currently ranks first on NL2Repo-Bench, QwenSWEBench, CoWorkBench, IFBench, and Agent's Last Exam (score). That said, these are vendor-published numbers — independent replication is still pending.\n\n## The Memory Bill\n\n[Yotta Labs](https:\u002F\u002Fwww.yottalabs.ai\u002Fpost\u002Fqwen-3-8-27b-specs-hardware-requirements-how-to-run-2026) runs the math: roughly 56GB at BF16 (H100\u002FH200 class), ~28GB at FP8 (48GB cards), and ~14-17GB at 4-bit quantization — Unsloth's local quantized build lands around 17GB, meaning a single 24GB RTX 4090 can run it. Compare that with fellow Chinese flagship Kimi K3: 2.8T parameters, self-hosting starting at 1.56TB of weight memory, cluster scale. That's the point of a 27B — not competing on size, but on deployability.\n\n## So What\n\nQwen3.8-Max goes API (\u002F per million tokens), 27B goes open — Alibaba has split the same generation's capability into two distribution tracks. Its predecessor Qwen3.6-27B was already one of the community's favorite local coding\u002Fagent models, and this generation adds multimodality and 262k context without raising the single-GPU barrier. For teams running their own inference, these are the most worthwhile new weights to eval this month. For everyone else, the answer is simple: the download is Apache 2.0 — whether it handles your workload, one test run will tell.","qwen3-8-27b-open-weights-release","2026-08-14T19:30:00Z","2026-08-14T21:05:30.391009Z","2026-08-14T21:05:30.391018Z",true,"agent",236,{"items":39},[40,45,50,55,60,65],{"id":41,"title":42,"news_slug":43,"published_at":44},"40095b51-97b0-4fd4-9b1d-f636c970572e","阿里 Qwen 团队发布 Qwen3.8-Max:2.4 万亿参数 MoE 模型首度开放权重","qwen3-8-max-2-4t-moe-open-weights","2026-08-07T02:00:00+00:00",{"id":46,"title":47,"news_slug":48,"published_at":49},"4c7f5330-3aff-458a-9ef5-f04cc5585703","微信视觉团队开源 WeMM 嵌入模型:2B 反超 8B 前基线,9B 达 MMEB-v2 80.6","wemm-embedding-wechat-multimodal","2026-08-26T21:07:30+00:00",{"id":51,"title":52,"news_slug":53,"published_at":54},"7ef479ae-66af-463a-802f-07a84ade93b1","商汤开源 SenseNova-U1.5-8B：原生多模态通吃生成编辑，短板全写进模型卡","sensenova-u1-5-8b-open-source-multimodal","2026-08-25T19:30:00+00:00",{"id":56,"title":57,"news_slug":58,"published_at":59},"96b989b7-992b-424e-a8c1-1568760150c1","小红书开源 dots3-note:280B MoE 多模态、512K 上下文,Apache 2.0 直接放行","dots3-note-preview-280b-open-weights","2026-08-18T23:10:00+00:00",{"id":61,"title":62,"news_slug":63,"published_at":64},"fb97a60d-69a1-4988-8de6-d1540ba63359","2.4B 参数读懂整页 A4:Cohere Labs 把最小的多模态模型挂上了 Apache 2.0","cohere-north-micro-vision-open-vlm","2026-08-18T13:30:00+00:00",{"id":66,"title":67,"news_slug":68,"published_at":69},"b1e41506-8ce0-4bbd-a11a-89d823998130","B 站 IndexTTS-2.5 开放权重:0.8B 参数零样本克隆五语种音色,8 维情感向量把情绪做成旋钮","indextts-2-5-bilibili-zero-shot-tts","2026-08-17T15:30:00+00:00"]