Locals runs many coding agents on a Mac, three times faster

Generated Computers has released Locals, a local model server made for AI agents rather than plain chat. The repository went up on 12 October. It had 432 stars 17 days later. The team describes itself only as Generated Computers.

The team says Locals runs several agents at a time on one Mac. The team also publishes numbers against the standard Mac stack.

What the published numbers show

The benchmarks on the README were measured on one setup. The Mac was a MacBook Pro with an M2 Pro and 16 GB of memory. The model was Qwen3.5-4B in 4-bit form. Generated Computers published every table.

With one request at a time, mlx-lm writes 61 tokens a second. Locals writes 64. The gap opens as more requests arrive. mlx-lm writes 87 tokens a second with four requests at once, and still 87 with eight. Locals writes 160 with four, 271 with eight, and 341 with sixteen.

The team also ran the LiveCodeBench benchmark on 40 programming problems, eight at a time. mlx-lm took 20 minutes. Locals took 8. That is 2.5 times sooner, on the authors' own measurement.

What that means for a Mac user

The target reader is a developer running Claude Code, OpenCode, Cline, Aider or OpenHands on their own Mac. One command opens any of them. For OpenCode on a project the command is locals code ~/my-project.

Locals keeps running in the background after the tool closes. Its port is 8321. The team says it installs next to Ollama, LM Studio or llama.cpp and uses the same model files. On an eight-agent task, the team's chart shows the time to the first word. It drops from 42 seconds through Ollama to 1.6 seconds through Locals.

The method behind the speed

Locals is built on a 2026 paper by Arip Asadulaev. The paper is called "Breaking the Space Barrier and its Application to Language Model Inference". It describes a JSON format or a tool call as a graph. Checking a reply against the format becomes finding a path in that graph.

Generated Computers turned the paper into an inference engine. Text the format already decides is written without running the model. The checks run on the graphics chip. A format takes about 3 MB of memory instead of up to 1.5 GB. The team says sixteen tool-calling agents finish 2.5 times sooner than with mlx-lm and llguidance.

Where it is not faster

The README names the cases where Locals is about the same as mlx-lm. One request alone, with nothing to reuse, is 64 against 61. Reading very long prompts is limited by the GPU's arithmetic. Models in formats outside the tuned kernels, such as the mxfp4 Mellum model, run at mlx-lm's speed. The M5's own neural accelerators cut the gain from three times to one and a half times.

The team asks for issues from M1, M3 and M4 users. Those chips have not been measured yet. The repo is Apache-2.0 licensed. A Mac with 16 GB of memory gives the best results on the default model.

Source

Locals by Generated Computers on GitHub.