llama-server Model Download Stuck? Router Status Checks and the b11401 Preview Fix

Find the Stuck Step Before Changing Versions

If a model keeps showing as downloading, do not immediately delete weights or start the same download again. Record the target model, incident time, and last meaningful log entry. Then determine whether the download is still progressing, loading has not finished, or the interface disagrees with the process state. This guide focuses on Router status notifications; the following checks do not require an upgrade first.

Run llama-server --version from the directory containing the actual executable. On Windows, use .\llama-server.exe --version and record the version and build output. Do not substitute -v, which enables verbose logging. If several installations exist, also save the full executable path so you do not inspect a new file while an old instance runs in the background.

A single baseline row is sufficient: executable path, version, launch time, port, model ID, and last observation time. Add only the status and new logs for later observations instead of rewriting that baseline. This keeps outputs tied to the correct launch when switching windows, restarting a test instance, or explaining the problem to someone else.

Check the Launch Command: Single Model or Router

Starting llama-server without specifying a model enters Router mode, which can use the cache, --models-dir, or --models-preset. Supplying an -m path starts an ordinary single-model instance. Check the actual command rather than guessing from a window title.

If you maintain two launch shortcuts, inspect the one involved and record its working directory, model source, and listening port. Direct all subsequent queries and log checks to that instance. If the failure reproduces only in ordinary single-model mode, investigate its own error instead of switching to Router just to apply this fix; changing architecture introduces another variable.

Also verify the address the frontend actually connects to. If the browser still uses an old port, healthy logs in a new window cannot explain the old page. Match the page, requests, and server window before comparing versions. When several models are active, isolate the one under investigation so another model's progress is not attributed to it.

Read Model Status and Download Events Without Changing State

Send GET /models to your own Router address, find the target by id, and inspect status.value. The values downloading, loading, and loaded represent different stages. A failure example is unloaded with failed:true and exit_code; do not look for a status value named failed. Do not add ?reload=1, which may unload models whose sources have changed.

A background download starting does not mean it has completed. The documented observation path uses download_finished or download_failed in /models/sse, followed by GET /models to confirm the list. Distinguish download completion from the model being loaded. Compare the same model at nearby times; joining the event stream partway through and seeing no completion event does not by itself prove download failure.

What b11401 Fixes—and What It Does Not Establish

Released on October 5, 2026, b11401 is a preview; the stable latest release remains v0.5.0. Its related PR29895 separates command and log transport for Router child processes, fixing interference with status notifications from log output.

This provides a specific lead when a process appears complete but Router remains in downloading status. Similar symptoms alone do not identify the cause, however. Preserve continuous logs around the incident, check explicit failures, and compare the operating mode and version. Ordinary single-model mode does not use this Router child-process communication path, and the update cannot rule out model-source, disk, or other independent problems.

When Testing the Preview, Keep a Recoverable Baseline

The following procedure makes comparison easier. Preserve the current program directory, launch arguments, and preset, and record model paths. Select the asset matching your system, architecture, and backend from the fixed b11401 release page, then extract it into a new directory. Do not mix executables and supporting libraries from old and new versions. Read the version output from the new directory before preparing the test launch; an archive filename alone does not establish the version running.

Choose an unused port for the test instance, retain the original mode and arguments for comparison, and make sure logs will not overwrite earlier records. Change only the program version, without simultaneously replacing the model or quantization, and do not let two instances compete over the same download destination. Observe whether the previously stuck step now progresses, keeping corresponding status, event, and error records. If results worsen, stop the test instance and restore the old directory and arguments. This is not an official automatic rollback guarantee or a claim that every model can be downgraded unconditionally.

Separate Windows Display Problems From Actual Failures

This update also changes ANSI color display in the Windows console, which is a separate symptom. The existing --log-colors off option disables colored logs and can help isolate display interference, but it is not a substitute for the Router pipe fix. If the only symptom is escape characters or incorrect colors, compare the text of the same log before judging the download result by appearance.

For a support request, prepare a minimal record: executable path and version, launch method, target model, times around the problem, and relevant output with credentials and private paths removed. State the exact step where the old version and preview each stop instead of treating a responsive window as success. This article is based on fixed source code and official documentation, without installation or runtime verification. If matching evidence is still absent, retain the existing setup until the cause is clearer before choosing another version.