Ollama 0.32.1 shipped on July 16 with two fixes that matter for running local models in agentic contexts. First: improved tool calling on Gemma 4, with more reliable tool-response continuations in multi-turn reasoning. If you use Gemma 4 with tools, this version reduces cases where the model loses track between one call and the next.

Second, critical: a recurring memory leak in the MLX model cache has been fixed, one that was increasing memory usage across requests. If you’ve noticed Ollama growing in RAM for no clear reason, this was it.

Why it matters to you. If you run local agents or workflows that call tools multiple times in the same session, this update changes system stability. The leak wasn’t just annoying: on repeated loads, it could push things into swap or crash the agent. The fix comes with better performance on MLX cache snapshots.

Two minor notes: MLX model loading now respects OLLAMA_LOAD_TIMEOUT, and agents with web search or fetch capabilities now clearly tell you to run ollama signin when authentication is needed, instead of failing silently. Anyone working with local agents knows how much clear error messages matter.

If you want to try it. Download v0.32.1 from the release page on GitHub (link in sources), and if you’d reported memory issues or unstable tool continuations on Gemma 4, test it on those same cases.

In detail

Context: Ollama as a local runtime for agents.

Ollama is the most widely used local runtime for running open-weight models on your own hardware, without sending prompts to an external API. Version 0.32, released a few days ago, introduced an agent interface and warnings for old models: the signal was clear, Ollama is positioning itself as infrastructure for self-hosted agents.

This 0.32.1 adds no new features: it fixes two problems that made 0.32 unstable in production.

The MLX memory leak: what caused it.

MLX is Apple’s framework for inference on Silicon. Ollama uses it when you run models on Mac. The leak documented in the issue linked to the release was recurring: the MLX model cache wasn’t freeing memory between requests, and every new request added allocation without releasing the previous one.

On an agent running for hours, this meant constant RAM growth until swap or crash. Version 0.32.1 fixes the leak and improves cache snapshot performance, which translates to faster model loads after the first run.

Gemma 4 and tool calling: the continuations problem.

Gemma 4 is one of the open-weight models Ollama supports for tool calling. The problem fixed here concerned multi-turn reasoning with tools: the model calls a tool, receives the response, and must continue reasoning by incorporating that response. In some cases, the continuation broke: the model lost context of the previous call or didn’t integrate the response in the next step.

The release note says “more reliable tool-response continuations,” and that’s the key phrase for anyone building agents: if your workflow involves multiple chained tool calls, this fix reduces cases where the agent “forgets” what it just did.

Other fixes.

OLLAMA_LOAD_TIMEOUT now works for MLX models too: before it was ignored, which caused unexpected timeouts on large models. And agents with web search or fetch capabilities now explicitly signal when ollama signin is needed, instead of failing with a generic error. Small details, but anyone who’s debugged an agent knows how much the right error message counts.

Implications for local agent users.

If you run agents on Ollama in contexts where stability and memory matter (for instance an agent monitoring feeds or processing files for hours), 0.32.1 is the upgrade that removes the risk of silent failure. It’s not a release that adds features: it’s a release that makes reliable what 0.32 promised.

For anyone evaluating whether to move an agent to self-hosted, this is good news: the most widely used local runtime is maturing on real agentic use cases, not just demos.