DeepSeek through 9Router: models and request checks

DeepSeek models can be accessed through a provider connection configured in 9Router. First distinguish a direct DeepSeek API account from a third-party service offering a DeepSeek model. Their keys, model names, limits, and charges can differ even when the displayed model family looks the same.

Start with one explicit provider and model. Add reasoning options or fallback only after a minimal request succeeds.

Direct API or another provider?

For direct access, follow DeepSeek’s first-call documentation and use credentials from your own DeepSeek account. The current documentation provides an OpenAI-compatible API at https://api.deepseek.com and also describes an Anthropic-compatible path. Use the protocol required by your client or gateway connection.

For another provider, use that service’s own endpoint and credential. A free allowance at a third-party service is not a free balance in a DeepSeek account. Likewise, a DeepSeek key cannot authenticate to an unrelated hosted gateway merely because it serves the same model family.

Discover names instead of freezing an old alias

Model aliases and available versions change. Select the model from the provider’s current catalog, then inspect the identifiers exposed by 9Router. Use that router identifier in the client. If you change the provider, repeat discovery; do not assume the previous ID maps to the same implementation.

Record the provider, model ID, and request settings alongside a test result. This matters when an alias is updated to a newer underlying model or a provider changes its deployment. The name alone may not explain a behavioral change.

Verify a plain request first

Check the local server and authenticated model list, then send a short text prompt without custom reasoning fields, tools, or a long history. If that succeeds, test streaming. Next add the particular feature your application needs, one at a time.

A useful coding check asks the model to perform one harmless operation against a test file. Inspect whether tool arguments are valid, whether the client executes the intended tool, and whether the result is incorporated into the following turn. A successful greeting does not establish agent compatibility.

Why does a thinking request return HTTP 400?

Reasoning fields are not interchangeable across providers. Upstream issue #5108 reports a 9Router translation problem involving DeepSeek-style thinking parameters and a strict OpenAI-compatible upstream. It is a version-specific report, not proof that all DeepSeek routes are broken.

Capture the installed router version, selected connection type, and exact error. Compare a minimal request with the failing reasoning request. Remove unsupported options, then consult the model/provider’s current parameter documentation. Do not repeatedly retry an unchanged invalid payload or assume that adding more token budget will repair a schema error.

Why does a long prompt fail below the listed context size?

A published context limit, a provider’s request-body limit, and the client’s budget calculation are different constraints. Issue #5097 reports a discrepancy between a catalog value and an OpenCode-served DeepSeek route’s observed rejection. The reported threshold belongs to that route and cannot be generalized to every DeepSeek provider.

Reduce the conversation and attachments, then compare the provider error with the router’s displayed limit. Preserve the error code and approximate input size without posting private content. A shorter prompt succeeding can help isolate a size problem, but it does not identify the exact maximum by itself.

Separate model quality from routing changes

If an answer changes after enabling a combo, inspect which backup actually answered. A fallback to another model family is not a DeepSeek-only comparison. If cost matters, check the serving provider’s usage records, including reasoning or cached-token billing where documented.

Keep one known working route as a baseline. Use fallback verification to test backups and OpenCode configuration if that is your client. This guide does not claim a live paid benchmark or guaranteed savings.