Ollama has released v0.32.4-rc0, a release candidate. It’s a maintenance version: no major new features, but fixes that matter for those running open-weight models on their own machines.

The most visible change is 8-bit LM head quantization. The LM head is the final layer of the model, the one that transforms internal numbers into words. Compressing it to 8 bits reduces memory usage without degrading text quality too much. If you have a GPU with limited VRAM, these megabytes are the difference between a model that loads and one that doesn’t fit.

Two stability fixes for those running local agents: a data race on the scheduler (two threads reading the same structure simultaneously, causing unpredictable behavior) and strengthened flaky tests that were generating false alarms. As we discussed in July when covering the memory leak fix in 0.32.1, long-session stability is the main frontier for Ollama users working with agents.

There’s also Laguna support on MLX: Poolside’s Laguna models, which we covered when they released the open-weight 118B, now run on Apple Silicon via Ollama. Plus a fix to expert quantization for Qwen 3.5.

Since this is a release candidate, if you use Ollama for daily work the advice is still to wait for the stable release. If you want to try Laguna on Mac, now’s the time to test.

In detail

What is the LM head and why quantization matters.

A language model is made of many stacked layers. The last one, called the LM head, takes the internal representation and projects it onto the vocabulary: for each possible word, a number indicating how likely it is as the next word. It’s a large layer, because a modern model’s vocabulary ranges from 30,000 to over 100,000 tokens.

When you quantize a model (compress it from 16 bit to 4 or 8 bit to make it fit in less memory), the LM head usually stays at full precision. It’s a conservative choice: the final layer is sensitive, and compressing it too much can visibly degrade text quality. Jesse Gross’s PR in this release quantizes the LM head to 8 bits, a compromise that halves that layer’s memory compared to 16 bits while keeping enough precision to not ruin the output.

For those with a GPU with 8 or 12 GB of VRAM, this might mean a model that didn’t fit before now loads. The exact benefit depends on the model: the final layer can weigh hundreds of megabytes in models with large vocabularies.

The data race fix.

A data race is a concurrent programming bug: two parts of the program read and write the same variable without coordinating. The outcome depends on timing: sometimes it works, sometimes it doesn’t, and reproducing the bug is hard.

Dhiltgen’s fix closes a data race on the scheduler loaded map, the data structure that tracks which models are in memory at any given time. If a local agent queries system status while another thread is loading or unloading a model, the map was read and written simultaneously. In most cases nothing happened, but in long sessions with multiple models the risk of unpredictable behavior was real.

MLX memory resident.

On Mac, Ollama uses MLX, Apple’s inference framework for Silicon. Before this change, when a model was unloaded to free memory, it had to be reloaded from scratch on next use. Dhiltgen’s PR keeps the model resident in memory even when inactive, reducing restart time. For those working with agents that switch models during a session (a small model for planning, a large one for execution), the time savings are concrete.

Laguna on MLX.

Poolside AI released Laguna S 2.1 as an open-weight model on July 23rd. Until now, running it on Mac required manual setup. Adding MLX support in Ollama means you can download and launch it with the usual command, no configuration needed. The model is a 118B parameter MoE: you need a Mac with generous unified memory, at least 64 GB to be comfortable with quantization.

The limits.

This is a release candidate, not a stable release. The fixes are in the changelog and PRs, but there’s no wide adoption yet to confirm no regressions. Numbers on memory saved by LM head quantization aren’t published: the exact benefit depends on the specific model. If you use it in production, wait for 0.32.4 stable.