Ollama Community Reports Windows Migration Defect; Model Names May Become Unreadable After 0.40.0 Update

A new compatibility issue has been reported with Ollama’s Windows model migration. A case filed on October 7 says that after an automatic update from 0.35.1 to 0.40.0, the previously usable qwen3.5:9b could no longer be accessed by its existing name: ollama show returned an error about a path containing an untrusted mount point, while /api/show and /api/chat returned HTTP 500. This is failure evidence provided by a single user; no maintainer has confirmed it so far. Issue report

The report narrows the failure to how model manifests are represented on disk. After the update, a local GGUF compatibility migration created named manifests under manifests-v2 as symbolic links pointing to blobs that were still present. The blob contained manifest entries for both the ggml and llamacpp execution backends; the user said the model could still be run using the llamacpp:<digest> name shown in the listing. This supports the assessment that reading the named manifest is blocked, but does not establish that all weights or migration results have been verified as sound. File attributes and logs

A symbolic link redirects a file path to another target, and an application must resolve that relationship when reading it. Microsoft’s documentation explains that Windows supports absolute and relative symbolic links, with relative paths resolved based on the link’s location. Therefore, a completed migration and an existing target file are not enough to prove that the serving process can successfully open the model by its original name. Microsoft technical documentation

The affected environment was Windows 11 Pro with Developer Mode enabled. The reporter suspects that permissions for creating links, combined with later read restrictions, caused the failure, and recommends that Windows use regular files instead. These remain unverified hypotheses, and no proposed fix has been attached to the issue. Issue status

For engineers maintaining local inference services, this case is a reminder that upgrade acceptance checks should cover existing model names and real API requests. Ollama’s official documentation confirms that Windows downloads updates automatically and that OLLAMA_MODELS can be used to specify the storage location. During troubleshooting, record the version, actual directory, link target, and process permissions. Next, watch to see whether upstream can reproduce the issue and whether a fix addresses both new migrations and links that have already been created. Ollama FAQ