大模型
dev.to ·
@Ammar Najjar
内存翻倍后，我的本地 AI 配置关注点变了
作者回顾在 16GB 统一内存 Mac 上运行本地 LLM 时，模型选择、量化和上下文长度都受内存限制。升级到 48GB 后，Qwen3.8-27B 等模型可用于日常，关注点转向模型在真实编码代理会话中的表现。当前配置为 LM Studio 搭配 Qwen3.8 与 Splash，OpenCode 提供代理式编码流程。作者指出上下文长度有直接成本，长会话会累积工具输出和代码，压缩可能提前打断流程，思考模式在部分任务上提升效果但增加延迟。
AI推荐理由 作者分享从 16GB 统一内存升级到 48GB 后本地 LLM 使用体验的变化，指出内存不再是主要瓶颈，注意力转向模型在真实编码代理会话中的表现。文中提到 LM Studio、Qwen3.8、Splash 与 OpenCode 的组合，并讨论上下文长度成本、压缩打断流程、思考模式带来的延迟等问题。属于第一手实践记录，对本地部署和编码代理使用者有参考价值，但为个人经验，缺少系统性基准数据。
62
正文
Six months ago, I started running local LLMs seriously on a Mac with 16 GB of unified memory. Most of my decisions were dictated by memory: which models fit, which quantization to use, and how much context I could afford.
Moving to 48 GB changed that.
Models such as Qwen3.8-27B became practical for everyday use, and my attention shifted toward a different problem: how well does the model work during a real coding-agent session?
Memory Was Only the First Constraint
My current setup uses LM Studio with Qwen3.8 and Splash, while OpenCode provides the agentic coding workflow.
With enough memory to run larger models comfortably, other constraints became much more visible.
Context length has a direct cost. Longer sessions accumulate tool output and code. Compaction can interrupt the flow earlier than expected. Thinking improves some tasks while adding latency to others.
A model that looks impressive in a short prompt can therefore feel quite different when it has to navigate a repository, inspect several files, use tools, make changes, and preserve enough context to continue reasoning about them.
That changed how I evaluate local models.
I care less about whether a model technically fits on my machine. I care more about how long I can work with it before the workflow starts getting in the way.
The Setup I Actually Use
After quite a bit of experimentation, the stack has become relatively small:
Mac with 48 GB unified memory → LM Studio → Qwen3.8 / Splash → OpenCode
The interesting optimization work now happens around that stack: choosing a sensible context size, deciding when thinking is worth enabling, understanding why compaction happens, and balancing model capability against interactive speed.
There is still plenty to experiment with, but having more memory has made the trade-offs much clearer.
I documented my current configuration, what changed since the original 16 GB setup, and the lessons from using it for actual development work in the full article:
https://ammar-najjar.com/blog/local-ai-setup-six-months-later/
If you use local LLMs with coding agents, I would be interested in comparing experiences. Once model size and memory stopped being your main constraint, what became the next bottleneck?