LM Studio and Qwen: How Local LLMs Handle Coding on a MacBook M4 Pro

Local language models can already be used to write and edit code without sending source code to the cloud, but how comfortable that is still depends heavily on the task and on how much hardware headroom you have. An experiment on a MacBook Pro with an M4 Pro and 48 GB of memory shows that LM Studio paired with modern open-weight models already delivers a solid result in chat mode, but with full-fledged agents it quickly runs into limits on memory, heat and execution time. The starting point is simple: cloud models are convenient, but they have limits, depend on the network, and, the main drawback for many developers, your data, code and prompts go to remote servers.

Running locally promises privacy and full control, but it requires understanding how a model eats RAM and VRAM, how much memory is left for context, and how formats such as GGUF and MLX differ. The tests ran on a MacBook Pro with an M4 Pro chip and 48 GB of unified memory, where the CPU and GPU share a single memory pool. That helps fit larger models, but it also means the model competes for resources with the IDE, Docker and dozens of browser tabs.

A separate part of the write-up covers choosing a model for your hardware. The author advises looking not only at size in billions of parameters, but also at specialization, quantization, function calling support and architecture type. For coding he took Qwen3-Coder 30B A3B Instruct in MLX and GGUF variants, and compared it with Qwen3-Coder Next, Qwen3.

5, Nvidia Nemotron-3 Nano and Gemma 4 26B A4B. The article explains the practical meaning of the abbreviations well: for example, A3B points to a MoE approach, where only part of the parameters of a large model are activated, so speed is closer to small models and quality closer to larger ones. LM Studio was chosen as the runtime: it makes it easy to download a model, start a local server, enable CORS and connect agents such as Claude Code, Open Code, Kilo Code and Aider.

The performance forecast for Qwen3-Coder promised about 150 tokens per second, but the actual measurement in LM Studio came in closer to 82 tokens per second, which immediately pulls the discussion from theory back to practice.

The most interesting part begins with the measurements. In plain chat mode, local models no longer look like a toy but like a working compromise. Qwen3-Coder 30B A3B Instruct in MLX 4bit completed the whole three-stage scenario in about 2 minutes 9 seconds and reached a final score of 8.

5 out of 10. Gemma 4 26B A4B in GGUF showed one of the best balances: about 2 minutes 23 seconds and a final 10 out of 10. The more reasoning-heavy models gave better results, but at a cost in time: Qwen3.

5 35B A3B reached 10 out of 10 in about 5 minutes 43 seconds, while Qwen3.5 27B stretched to almost half an hour. The conclusion from this part is sober: in response speed, local models are in places already comparable to cloud ones, especially without thinking mode, but in the same time they more often fall short of them in quality.

Recent MoE models, however, look noticeably more practical than dense ones.

In agent mode the picture changes sharply. Context grows, the number of calls increases, and seconds turn into minutes or even tens of minutes. Aider with the same Qwen3-Coder MLX 4bit finished the scenario in 2 minutes 50 seconds with a score of 9.

5, Open Code took 7 minutes 33 seconds with a score of 9, but Kilo Code with the same model took 15 minutes 5 seconds and reached only 6 points. With the heavier Qwen3.5 35B A3B, Kilo Code already ran for 57 minutes 3 seconds, although the final quality rose to 9 out of 10.

Claude Code with Gemma 4 26B finished the experiment with the maximum score of 10 out of 10, but spent 21 minutes 14 seconds in total, and the Claude Code and Qwen3-Coder pairing crashed outright for lack of memory for context. Meanwhile the laptop visibly suffered: the GPU heated to about 100 degrees, the fans barely went quiet, and swap in some scenarios ballooned to 20 GB. Against that backdrop, cloud agents were simply more convenient: for example, Kilo Code with Qwen3.

5 Plus gave 9 out of 10 in 6 minutes 53 seconds, and Claude Opus 4.6 gave 10 out of 10 in 12 minutes 15 seconds, though for money.

The bottom line is simple: local LLMs can already be seriously considered for private chat, one-off refactoring tasks and simple scenarios where control over data matters more than absolute speed. But if you need a permanent agent mode on a work laptop, especially alongside an IDE, a browser and Docker, the local stack remains a compromise for now. The most sensible scenario from this experience is to take recent MoE models, use simpler agents such as Aider or Open Code, and, where possible, move the local model to a separate machine such as a Mac mini.