Hacker News 中文摘要

RSS订阅

Qwen3.8-27B -- Qwen3.8-27B

文章摘要

阿里通义千问发布Qwen3.8系列模型,包括27B参数的多模态稠密模型,性能超越Qwen3.7-Plus,在编程和办公场景表现优异,支持262K原生上下文并可扩展至百万token,同时开源了Max级模型。

文章总结

阿里巴巴通义千问团队宣布,兑现此前关于开源Qwen3.8模型权重的承诺。新发布的Qwen3.8-27B是一款原生多模态稠密模型,仅用270亿参数便在整体性能上超越Qwen3.7-Plus,尤其在真实编码和办公工作流中表现突出。该模型支持26.2万原生上下文,并可通过YaRN技术轻松扩展至100万token。此外,团队近期还开源了Qwen3.8-2.4T-A95B(Max级别)的权重。无论是本地部署轻量级应用,还是构建智能体,开发者均可下载使用。模型已上线Hugging Face和ModelScope平台。

评论总结

根据评论内容,主要围绕Qwen 3.8 27B模型展开,观点如下:

1. 性能显著提升,可本地运行 - 多数评论认可模型在27B参数下达到Opus 4.6级别性能,且可在高端个人电脑上运行。 - 关键引用:chvid "These are massive improvements - and something you can actually run on a laptop.";NorwegianDude "we now have a local model that is runnable on a high-end personal PC that trades blows with the leading model Claude Opus 4.6 Max."

2. 基准测试表现亮眼,但需谨慎 - 模型在DeepSWE等基准上超越Opus 4.6 Max(42.2 vs 40),但部分评论提醒基准测试不等于实际使用体验。 - 关键引用:scrlk "Beats Opus 4.7 Max (w/ Claude Code) on DeepSWE (42.2 vs 40).";mickeyp "Model benchmarks are useful, to a point... one open weight model might 'meet' Opus on benchmarks, but then fail to follow a simple answer format."

3. 对模型变体和生态的期待 - 用户希望推出更小(如10B)或MoE变体(如35B A3B),并关注量化版本(如1bit)的可用性。 - 关键引用:tosh "I hope there also will be a new ~10b variant";KronisLV "I hope really badly that we'll get a new 35B A3B or similar MoE model!";jedbrooke "I hope the bonsai team makes another 1bit quant of this model."

4. 实际使用与成本考量 - 部分用户关注本地运行速度(如4090上约48 tps),并对比API成本(如DeepSeek、Gemini等)。 - 关键引用:kristopolous "q4km is about 48 tps on a 4090.";ramon156 "Opus might be a slight bit better... but it's also extremely expensive, and I hit the 5 hour limit way too quick."

5. 多模态与未来潜力 - 模型支持多模态,用户期待与工具(如MCP)结合,并认为本地模型是未来方向。 - 关键引用:tosh "also cool: Qwen 3.8 27b is multi modal!";brcmthrowaway "This with ddg mcp to fill in world knowledge. Are local models the future when computer architectures catch up?"