DeepSeek through 9Router: models and request checks

For a direct DeepSeek account, add an OpenAI-compatible provider in 9Router with base URL https://api.deepseek.com, save your DeepSeek key, and expose the chosen model under a local prefix. A third-party host serving DeepSeek models has its own endpoint, key and billing; this recipe is specifically for DeepSeek’s API.

You need a working local router, an eligible funded DeepSeek API account, curl and jq for the Bash/zsh examples. Each chat test may consume paid usage. The instructions below are source-reviewed; this Wiki has not run a paid DeepSeek benchmark.

1. Establish the direct API baseline

Set DEEPSEEK_API_KEY in your local environment. As checked on October 23, 2026, DeepSeek’s first-call documentation lists deepseek-flash; check its current catalog before copying that alias into an older installation. Start with thinking explicitly disabled so the comparison has fewer moving parts.

curl --fail-with-body -i https://api.deepseek.com/chat/completions \
  -H "Authorization: Bearer $DEEPSEEK_API_KEY" \
  -H 'Content-Type: application/json' \
  -d '{"model":"deepseek-flash","messages":[{"role":"user","content":"Return a short greeting."}],"thinking":{"type":"disabled"},"stream":false}'

A successful baseline is a complete chat response with choices[0].message.content. If this fails, fix the direct account or payload first. Do not add 9Router to a request whose upstream credential already fails.

2. Save the DeepSeek route

Open Providers → Custom Providers (OpenAI/Anthropic Compatible) → Add OpenAI Compatible.

Setting	Enter
Name	DeepSeek direct
Prefix	ds-local (an unused local namespace)
API Type	Chat Completions
Base URL	https://api.deepseek.com
Optional check key/model	DeepSeek key / the exact upstream model ID

Select the API type first, then the URL. Although the form’s hint mentions /v1, the reviewed executor appends /chat/completions to this base, matching DeepSeek’s documented route. Do not substitute the separate /anthropic endpoint into this Chat Completions recipe.

Check may make an inference request when model discovery fails. Its API Key (for Check) is temporary validation input. Click Create, open the node, then Add API Key: give the connection a name, enter your DeepSeek API Key, set Default Model to deepseek-flash or your verified upstream choice, and Save.

Use Import from /models to discover models, or type the exact upstream ID into the models field and click Add. Copy the resulting model row; with the example prefix and model it is ds-local/deepseek-flash. The upstream Default Model remains unprefixed.

3. Repeat through the router

Set NINEROUTER_API_KEY to the local router key and ROUTER_MODEL to the copied local ID. Keep the same prompt and thinking setting as the direct baseline.

curl --fail-with-body --noproxy '*' http://localhost:20128/v1/models \
  -H "Authorization: Bearer $NINEROUTER_API_KEY"
jq -n --arg model "$ROUTER_MODEL" \
  '{model:$model,messages:[{role:"user",content:"Return a short greeting."}],thinking:{type:"disabled"},stream:false}' \
  > /tmp/deepseek-router-probe.json
curl --fail-with-body --noproxy '*' -i \
  http://localhost:20128/v1/chat/completions \
  -H "Authorization: Bearer $NINEROUTER_API_KEY" \
  -H 'Content-Type: application/json' \
  --data-binary @/tmp/deepseek-router-probe.json

Check both the returned chat body and the selected provider in the router log. The model-list response is discovery evidence only. If direct succeeds but local fails, compare the exact model ID, saved key, base URL and request body before adding a combo.

4. Add thinking as a separate comparison

DeepSeek’s thinking documentation describes an enabled thinking mode and reasoning effort. For the supported direct model, replace the disabled field with:

{"thinking":{"type":"enabled"},"reasoning_effort":"high"}

Merge those fields into the existing request object; this fragment is not a complete chat request. Run the direct version first, then the local version with the identical fields. Expect reasoning-capable responses to differ from the disabled baseline; do not assume another host accepts DeepSeek’s parameter names. With the Python OpenAI SDK, DeepSeek documents putting thinking inside extra_body.

Issue #5108 reports a translation failure involving thinking fields and a strict upstream. If only the enabled local request returns 400, keep the working disabled route, record the router version and redacted error, and inspect the installed release’s translation support. More retries or a larger token budget cannot repair an invalid schema.

Context and tools need their own checks

For a long-context rejection, shorten the same request until it succeeds, retaining the provider error and approximate size. A catalog context figure is not proof of the route’s effective limit; issue #5097 documents a route-specific discrepancy. Do not generalize its threshold to all DeepSeek accounts.

Before using an agent, test one harmless tool operation in a scratch directory and verify the tool result reaches the next turn. If enabling fallback changes the serving model, label the result accordingly. Continue with OpenCode setup or proxy diagnostics once the plain route is established.