Managed-model output budgets#

Starting with rc.76, “MATLOOK Online Models” synchronizes the platform's verified output capacity, per-request budget, output parameter and reasoning configuration. Policy refreshes retain these platform settings. Quit normally, upgrade the client and reopen the conversation; repeated manual budget edits to the managed connection are unnecessary.

Context capacity and per-request output budgets differ. Thinking, answers and tool arguments share the output allowance. A large context does not imply equally large file output in one request. This deployment configures GLM-5.3-Flash-GGUF for 32,768 output tokens per request and leaves reasoning at the upstream default. Other models retain their individually verified settings.

Reaching the configured ceiling can still truncate output. MATLOOK blocks incomplete tool arguments and retains the error; split large files before retrying. A budget refresh never replays a previously failed tool call and does not change offline models or local data management.

Model connection recovery#

After a client disconnects, MATLOOK attempts to collect the original response to establish completion. Historical unknown requests stay recorded and protected against replay. After a cooldown, the connection admits one independent recovery probe. Meaningful output with normal protocol completion restores configured concurrency without certifying historical requests as completed. New unknown work entering quarantine enables recovery protection again.

Administrators can inspect configured concurrency, capacity limit, active requests, uncertain requests and historical records in the connection occupancy panel. Historical records alone no longer keep a recovered connection at single concurrency. Reconcile an old record only after verifying that its upstream work has stopped.

Before you begin#

Obtain the connection details and permission to use your chosen service. Models differ in text, vision, tool and context capabilities. Check the provider's current service availability and pricing.

Configure a desktop connection#

Open Settings → AI and add a supported connection type.
Supply the service address, model ID and authentication details required by that connection. Use the provider's complete Base URL for compatible endpoints.
Save, inspect available models and test a simple question without confidential project information.
Set the default if needed, then check the model selected in the conversation composer.

Defaults versus existing conversations#

Changing a default for new conversations may not update existing conversations. A remotely controlled conversation uses connections available on its execution computer. Some web connections are managed by an administrator, so regular users may not have key configuration controls.

Utility requests and fallback models#

Titles, summaries and utility analysis without an explicit model follow the current conversation model, unless the connection has a valid utility-model override. An explicitly specified tool model remains independent; omit the tool's model parameter to keep following the current utility policy. Changing the conversation model updates subsequent utility defaults.

If a managed online model fails or has no available execution slot, an authorized alternate may handle that request. The conversation displays a separate notice identifying the selected model, actual model and reason. This does not permanently change the model selected in the composer. The same transition is shown once per turn. Background title and summary fallback does not add chat notices, and notices are kept out of assistant text, tool results and conversation titles.

The fix requires desktop version 0.5.0-rc.66 or later; precise fallback reasons also require the matching server update. Existing affected titles can be renamed in the app. Local utility requests stay on their configured connection and do not automatically switch to an online provider. A stale utility default may recover to the current conversation model on the same connection; an unavailable explicitly requested tool model returns an error.

Troubleshoot no response#

If the upstream model account has insufficient credit, contact an administrator or select another authorized model. This is separate from your available MATLOOK quota and should not trigger repeated retries. Before reconciling an execution whose status is unknown, administrators must verify that upstream generation and queued work have ended, so the same work is not executed twice.

Check whether the target is online or waiting for approval, then verify the model ID, reachable endpoint, credential permissions and service allowance. Test a minimal text request before adding attachments. Never paste keys or credential-bearing URLs into chat or screenshots.

Next step#

See local models for local inference. A connection enables inference; sources and tools remain subject to permissions.