Ollama 0.32.1 fixes a memory leak affecting those running local agents in the background and improves tool calling in multi-turn reasoning. It’s a maintenance release: no new features, just two fixes that stabilize what used to break down during long sessions.

As we covered when 0.32.0 landed, Ollama has focused on the agent experience: you run a command and end up in a chat that searches the web, writes code, and delegates tasks. Version 0.32.1 refines exactly how that tool mode works.

The memory leak was the serious problem. If you left an agent running in the background for extended sessions, memory usage would grow without being released. Anyone using open-weight models for recurring tasks (like an agent processing mail every morning or a mini-radar aggregating RSS feeds) found themselves with processes that slowed down until they stopped. The fix resolves this behavior.

Improved tool calling applies to multi-turn reasoning: when the agent calls a tool, reads the result, and decides the next call based on that output. Previously it could lose track or repeat the same empty call. Now the sequence holds up better.

Why it matters to you. If you’re running Ollama for local agents, this is an upgrade to make right away. It doesn’t change your workflow, but it makes reliable what used to break after thirty minutes. If you’re not using local agents yet, this release isn’t what gets you started: the value is in stability, not novelty.

The source is the official changelog on GitHub, published July 11, 2026.

In detail

What came before.

Version 0.32.0, released July 11, 2026, changed how Ollama presents itself. Before, you’d run ollama and get a shell to download and run models. Now you run the same command and land in an agent interface offering chat, coding, and task delegation, with web search built in. The default model for this experience is glm-5.2:cloud.

The 0.32.0 release also introduced a deprecation warning for older agentic models: CodeLlama, Qwen2.5 (including the coder variant), Llama 3.x, Mistral, StarCoder, and base tags of DeepSeek-R1. If you try to run one with the agent command, Ollama warns you first. It’s a clear signal: local tooling is pivoting toward models designed for agents, not generic ones trained years ago.

What changes in 0.32.1.

Version 0.32.1 is a fixes release addressing two concrete points in the agent experience introduced in the previous version.

The memory leak showed up when an agent stayed active in the background for extended sessions. Memory allocated to manage context, tool results, and conversation state wasn’t released between turns. The process would grow in RAM until it saturated machine resources, causing progressive slowdowns and, in worst cases, crashes. For anyone running Ollama on a laptop with 16 or 32 GB of RAM, where the model itself already occupies most available memory, the problem surfaced after 20-30 minutes of agent activity.

The fix corrects the lifecycle management of context objects, releasing memory between turns when no longer needed.

The multi-turn tool calling addresses the agent’s ability to chain calls to external tools. In a typical flow, the agent calls a search tool, reads the result, decides whether to call another tool or respond. Before 0.32.1, in long sequences it could lose track of state or repeat already-made calls. The improvement touches the routing logic between tool results and the next reasoning step.

How much to trust it.

The changelog is verifiable on Ollama’s official GitHub repository. The 0.32.0 release notes are public and detailed. For 0.32.1, the described fixes align with issues reported by the community in the weeks before, but the specific 0.32.1 release notes should be checked directly on the corresponding tag.

There are no numerical benchmarks tied to the tool calling improvement: it’s a stability fix, not a measurable performance leap.

Practical implications.

If you have local agents running on a schedule (morning email triage, feed digest, page monitor), 0.32.1 makes them reliable over long sessions without manual restarts. If you only use Ollama for short interactive chat, the memory leak probably didn’t hit you, and the upgrade is optional but recommended for safety.

For building stable local agents, the playbook on how to delegate a multi-step task to an agent still holds: Ollama’s fix addresses the infrastructure, but result quality always depends on how you write the brief.