Ollama quietly made Macs a better place to run open models
The 0.32.4 releases add Laguna support on Apple GPUs via MLX, smarter mixed quantization for mixture-of-experts models, and a fix for Qwen3 MoE decoding. Small, unglamorous, and exactly the sort of update that decides whether local AI is usable.
While the week’s attention went to a 1.4-terabyte model release, the tool most people actually use to run AI on their own machine shipped a set of changes that matter more for everyday use. Ollama’s 0.32.4 line adds support for Poolside’s Laguna models, specifically the XS 2, XS 2.1 and S 2.1 variants, on Apple GPUs through MLX, Apple’s own machine-learning framework for its silicon.
The rest of the release notes read like plumbing, which is the point. The update reworks how quantization is applied to mixture-of-experts models. Quantization means storing a model’s numbers at lower precision so it fits in less memory, a bit like saving a photo as a smaller JPEG. Do it bluntly and quality drops; do it selectively and you get most of the size saving for almost none of the loss. Ollama now reads the source model’s own configuration and applies one coherent policy across both the dense layers and the routed expert layers, keeping the router and the tied output head at full precision, quantizing attention and expert projections, and selectively promoting the sensitive expert down-projections back up. It also fixes decoding for Qwen3 mixture-of-experts models whose experts were quantized differently from one another, and reports roughly 4 to 9 percent faster packed gate and up-projection performance on an M5 Max.
Here is why unglamorous work like this decides whether local AI is real. Running a model on your own laptop is almost never limited by intelligence; it is limited by memory. The gap between a model that fits in your machine’s RAM and one that does not is the gap between two seconds and two minutes per answer, or between running and not running at all. Every incremental improvement in how carefully a model gets compressed moves a class of models from the second category to the first. The mixture-of-experts angle matters especially because that architecture is now standard at the top end, and MoE models were historically awkward to quantize well: the router that decides which expert handles what is tiny but extremely sensitive to precision loss, so treating it like every other layer wrecks the model.
What this means for you: If you have never run a model locally, the short version is that you can, that it is free, that nothing you type leaves your computer, and that a current Mac with 16 GB or more is a decent machine for it. Ollama is a single install and one command. If you already run models locally on Apple silicon, this release is worth taking: Laguna on MLX is a genuine addition rather than a rename, and the MoE quantization changes may let you fit a model you previously had to skip. If you are on Windows or Linux with an Nvidia card, the MLX parts do not apply to you, though the Qwen3 fix does. Either way, this is the kind of steady, boring progress that quietly widens what runs on hardware you already own.
Sources
Source: https://github.com/ollama/ollama/releases/tag/v0.32.4
OpenAI looked at 800,000 work chats and found people quietly doing other people's jobs
A new OpenAI study says 43.5 percent of job-specific ChatGPT use involves tasks that traditionally belong to a different profession. Marketers troubleshoot code, salespeople run their own data analysis, and small companies do it most of all.
