Your Ollama Models Go Stale. OllaMan 3.8 Tells You When.

You pull a model once, it works, and you never think about it again. Meanwhile the publisher keeps working on it. They fix a chat template, re-quantize the weights, repackage the layers, and push the result under the same name and tag. gemma4:latest on ollama.com today is not necessarily the gemma4:latest on your disk.

Ollama has no way to tell you this. ollama list shows when you pulled a model, not whether the registry has moved on. To find out, you would have to re-pull every model and see what downloads, which nobody does.

OllaMan 3.8 checks for you. Every installed model gets compared against its registry, the ones with a newer version get an icon, and one click brings them up to date.

A real example

While writing this post, we opened OllaMan on one of our own machines. Four of its five models had a newer version waiting.

The most striking was gemma4:latest, pulled five months ago. Locally it is a single 8.95 GB weights layer. The version on ollama.com today is a different build: the weights are split into separate model, projector, and draft layers, about 6.1 GB in total. Same name, same tag, noticeably different files. Without a check, there was no sign of it.

Spotting updates

Open Installed in the sidebar. A model with a newer version shows an update icon, an arrow in a circle, right below its name.

In the screenshot, four models have the icon. The fifth, a Hugging Face GGUF, matches its registry and shows nothing. No icon means up to date.

You don't need to start the check yourself. OllaMan runs it when you open Installed and the last check is more than 6 hours old, and in the background when results are more than 24 hours old. Checks are quiet. If you're offline, nothing breaks; OllaMan just tries again next time.

Update detection is free for everyone.

One click to update

Click the update icon and OllaMan pulls the new version. You can also open the model's ⋯ menu and choose Update model.

The update goes through the regular download queue, so progress shows on the Downloads page like any other pull, and you can update several models at once. Ollama only downloads the layers that changed. When a publisher only edits the template or parameters, the update finishes in seconds rather than re-downloading gigabytes of weights. When the icon disappears, you're on the new version.

Applying updates is an OllaMan Pro feature. Without Pro, clicking the icon shows an upgrade prompt; detection keeps working either way.

Skip a version you don't want

Not every update is welcome. Maybe the old build works well for your prompts and you'd rather not re-test. Open the ⋯ menu and choose Skip this version. The icon goes away and stays away until the publisher releases another version after that one.

How OllaMan knows a model changed

Every model in Ollama is described by a manifest, and the manifest has a digest: a SHA-256 fingerprint of exactly which layers the model consists of. If the publisher changes anything, the digest changes.

OllaMan compares the digest of your local manifest with the one on the registry:

• ollama.com models (llama3.2, qwen3:8b, user/model:tag): a HEAD request for the manifest, reading the ollama-content-digest header. Nothing is downloaded.
• Hugging Face GGUF models (hf.co/org/repo:quant): a GET for the manifest, hashed with SHA-256.

Same digest means up to date. A different digest means an update. If the registry answers 404, 401, or 403, the model only exists locally (you created or imported it), so there's nothing to compare against. Cloud models (:cloud) are skipped, since they don't run on your machine anyway.

Up to four models are checked at a time, and results are cached, so opening Installed doesn't hammer the registry. The Update Models guide has the full list of supported model sources and a short FAQ.

Also new in 3.8

Model updates are one part of v3.8.0. The release also brought:

• Web search in chat. Agents get web_search and web_fetch tools, switched on per agent with a Search toggle in the composer. Use the built-in search with no setup, or plug in Ollama web search, SearXNG, or Tavily. See Web Search.
• Context length and usage. Set num_ctx per agent or per session, capped at what the model was trained for. A small ring next to the composer shows how much of the window each reply used and warns you before older messages get cut off. See Context Length & Usage.
• Export chats. Save any conversation as Markdown to read or share, or as JSON to keep everything, images included. See Export Chats.

Updating OllaMan

Existing installs get an auto-update prompt. To get it now, download from ollaman.com. Then open Installed and see how many of your models have quietly moved on.

---