AMAZINGINDEX.COM 每日 AI 简报
53.0
VOL. 2026.06
2026.06.16
← 返回 2026.06.16 日报
日报快照 · Daily Snapshot
NO. 012

本地35B模型替代Claude日常编码

#ARTICLE HackerNews 2026.06.16
推荐指数 60.0 NO. 012 · 2026.06.16
发布2026/06/15Score413Comments231

HN用户用量化版Qwen3.6 35B(仅3B活跃参数)在Mac Studio/MacBook上完全离线完成Django+Wagtail网站重构。证明消费级硬件+量化模型已能支撑实际开发工作流,隐私敏感者和离线场景可参考。

本地35B模型替代Claude日常编码

这个案例的关键不是模型能力,而是参数激活策略——35B总参数只跑3B活跃参数,用MoE架构的稀疏性换速度,这是本地部署能用的核心前提。之前大家试本地模型觉得慢,往往是全参数加载,没走这条路径。

Wagtail这个细节更值得注意:小众框架没有大量训练语料,模型仍能完成任务,说明当前代码模型的泛化能力被低估,不是只能写React和Python标准库。如果你在评估本地模型能不能覆盖内部技术栈,这个信号比benchmark更有参考价值。

真正门槛其实是上下文长度和工具调用稳定性,帖子没提这部分。建议想复现的人先测长文件编辑(>2000行)和连续多轮agent执行的成功率,这是日常编码和demo的区别。

意见分歧 216 条评论

核心争论:本地量化模型能否真正替代云端大模型用于日常编码,速度、成本与稳定性仍存争议

tumetab1

Not yet, tried Gemma 4 on an Apple M4 but the tok/s is significant lower than the cloud offering. Also,the lack of enterprise tooling to help selected an appropriate model and tooling to run a local LLM does not help.

arjie

Not “local” and not interactive coding but sharing since it might be helpful. I have 2x RTX Pro 6000 Blackwell running DeepSeek V4 Flash. I get 160 tok/s raw but it’s a reasoning model. For my use case, I have it auto-write code and another system auto-review the code. I occasionally use it wit

leptons

Have you measured your electricity consumption for this rig? I have to wonder how much it would cost you per month.

替代方案: ClaudeClaude CodeCodexCursorGemma 4DeepSeek V4 FlashKimi 2.6GLM 5.1OpenRouterFireworksQwen 3.6 27b denseClaude Haiku 4.5Claude Sonnet
查看原文 →