What local AI coding model can a contractor run on a Mac Studio or DGX Spark?

Question

Accepted Answer

Recursive Flint, a new open-weights coding model released Oct 21, 2026, is built to run on your own hardware. Recursive says it is a mixture-of-experts model with 40B total and about 2.5B active parameters. It's free under the Apache 2.0 license and published on Hugging Face in two local builds: MLX INT4 weights for Apple Silicon and an NVFP4 GGUF for NVIDIA. Mac Studio: the INT4 Mac build is about 22.6 GB, so it fits a Mac Studio with 48 GB or more of memory. WIRED's hands-on was on a 48 GB Mac. DGX Spark: the GGUF model card recommends 48 GB of memory, and the DGX Spark we list has 128 GB of unified memory, shared with DGX OS. Recursive says Flint runs at 125+ tokens per second on an M4 Pro MacBook Pro and 150+ on a 64 GB RTX Spark at short contexts. Those are Recursive's own numbers, and so are all of its benchmark scores. We haven't seen independent tests yet, or speeds at long contexts. What it's for: Flint is a coding model. Use it to build small tools, scripts and spreadsheet automations for your shop, not to write bids or proposals. Recursive points Mac users to its own Rec launcher and llama.cpp. oMLX, the MLX model server for Macs, merged native Flint support on Oct 23, 2026. It's in oMLX's 0.7.3 development builds (v0.7.3.dev1, Oct 24, and later), but not yet in a stable release. DIY-first approach: download the build for your machine from Recursive's Hugging Face page, try it on one real task like a script that renames job photos, and check the output before you rely on it. For setup help see our Mac Studio guide, DGX Spark setup guide and model and hardware fit guide, plus the Mac Studio listing and our local AI category.

What local AI coding model can a contractor run on a Mac Studio or DGX Spark?

Sources