Ask HN: Who uses open LLMs and coding assistants locally? Share setup and laptop

What's your tokens/s on that?

> Which model(s) are you running (e.g., Ollama, LM Studio, or others)

I'm running mainly GPT-OSS-120b/20b depending on the task, Magistral for multimodal stuff and some smaller models I've fine-tuned myself for specific tasks..

All the software is implemented by myself, but I started out with basically calling out to llama.cpp, as it was the simplest and fastest option that let me integrate it into my own software without requiring a GUI.

I use Codex and Claude Code from time to time to do some mindless work too, Codex hooked up to my local GPT-OSS-120b while Claude Code uses Sonnet.

> What laptop hardware do you have (CPU, GPU/NPU, memory, whether discrete GPU or integrated, OS) and how it performs for your workflow?

Desktop, Ryzen 9 5950X, 128GB of RAM, RTX Pro 6000 Blackwell (96GB VRAM), performs very well and I can run most of the models I use daily all together, unless I want really large context then just GPT-OSS-120B + max context, ends up taking ~70GB of VRAM.

> What kinds of tasks you use it for (code completion, refactoring, debugging, code review) and how reliable it is (what works well / where it falls short).

Almost anything and everything, but mostly coding. But then general questions, researching topics, troubleshooting issues with my local infrastructure, troubleshooting things happening in my other hobbies and a bunch of other stuff. As long as you give the local LLM access to a search tool (I use YaCy + my own adapter), local models works better for me than the hosted models, mainly because of the speed and I have better control over the inference.

It does fall short on really complicated stuff. Right now I'm trying to do CUDA programming, creating a fused MoE kernel for inference in Rust, and it's a bit tricky as there are a lot of moving parts and I don't understand the subject 100%, and when you get to that point, it's a bit hit or miss. You really need to have a proper understanding of what you use the LLM for, otherwise it breaks down quickly. Divide and conquer as always helps a lot.

Running local LLMs on laptops still feels like early days, but it’s great to see how fast everyone’s improving and sharing real setups.

I keep mine pretty simple: my desktop at home has an AMD 7900XT with 20gb VRAM. I use Ollama to run local models and point Zed's AI integration at it. Right now I'm mostly running Devstral 24b or an older Qwen 2.5 Coder 14b. Looking at it, I might be able to squeak by running Qwen 3 Coder 30b, so I might give it a try to test it out.

On a Macbook pro 64GB I use Qwen3-Coder-30B-A3B Q4 quant with llama.cpp.

For VSCode I use continue.dev as it allows to set my own (short) system prompt. I get around 50token/sec generation and prompt processing 550t/s.

When giving well defined small tasks, it is as good as any frontier model.

I like the speed and low latency and the availability while on the plane/train or off-grid.

Also decent FIM with the llama.cpp VSCode plugin.

If I need more intelligence my personal favourites are Claude and Deepseek via API.