Use ChineseModels in Codex and Kimi Code

You can now follow dedicated ChineseModels guides to add available text models to Codex and Kimi Code. Each guide includes a prompt you can paste into Codex, a downloadable Windows installer, model-refresh instructions, and the limits of the integration.

Start with the Codex guide, the Kimi Code guide, or the focused DeepSeek in Codex guide. On Model Square, text-model cards also link to setup instructions with that model already selected in the prompt.

Discover models with your own key

The installers fetch the authenticated /v1/models endpoint. They preserve exact model IDs and exclude image and video generators from the text picker, including known video families whose metadata may also contain a chat endpoint label.

Our October 22, 2026 check found 65 text/chat candidates and excluded 28 generation models. That is a dated account-level snapshot, not a promise that every account will always see the same list. Rerun the refresh step to pick up catalog changes.

Discovery tells you what an account can select. It does not certify that every model supports every coding-agent tool or thinking option.

Codex: a local adapter and a separate profile

The tested Codex setup uses a local authenticated adapter to translate its Responses requests into ChineseModels Chat Completions requests. The Windows installer stores the gateway key with Windows DPAPI and prepares a separate configuration profile.

Launch it with codex --no-daemon --profile chinesemodels, then use /model. The setup prompt preserves your main OpenAI configuration so Astra and other OpenAI models remain available in normal Codex sessions. This tested configuration uses one provider per session. Replacing the main provider and catalog also replaces the normal picker; the installer therefore keeps the two setups separate.

The adapter maps reasoning settings by model family. DeepSeek V4 has off, low, high and max choices; Kimi K3 has low, high and max. Some families always think, while others expose an on/off toggle. A “high” label on a toggle-only model means on, not a distinct depth tier.

The adapter supports text and tool-message translation. It emits completed streaming output in batches and has a conservative 32,768-token context setting. It does not implement vision, generation endpoints, hosted search or remote response compaction. The guide explains those limits before installation.

Kimi Code: retain your existing Kimi account

The Kimi installer adds a native OpenAI-compatible provider and model aliases alongside existing models. It preserves your Kimi login, default model, hooks and other providers. Once you restart Kimi, use /model or launch an exact alias:

kimi -m "chinesemodels/DeepSeek-V4-Flash"

Kimi's native credential field stores the API key in its local configuration. The guide also explains the environment-variable alternative.

Thinking behavior deserves a separate check. In Kimi Code 0.29.1, our native wire test forwarded the chosen standard reasoning effort but did not send each vendor's custom thinking-toggle fields. A toggle in the interface therefore does not prove upstream reasoning can be switched off. The guide distinguishes model metadata from verified request behavior.

What we tested

Codex's model listing returned all 65 candidates. Local adapter checks covered 165 model/effort combinations, tool-name translation, authentication and reasoning replay. Representative live text/thinking calls succeeded across DeepSeek, Kimi, GLM, MiniMax, Qwen, Hy3 and Doubao. A DeepSeek tool-result round trip passed.

Kimi accepted the configuration and completed a live DeepSeek-V4-Flash reply. That is a connectivity check, not certification of all 65 models for coding tasks.

Before using a different model for a large project, test a short reply, a supported thinking setting, and a tool-result round trip. Keep your configuration backup so you can restore the previous setup.

Copy a prompt and get started

Choose your client guide, copy the setup prompt, and enter your API key through the installer's local hidden prompt. Never put a customer key into a public issue, source file or browser URL.

• Set up Codex
• Set up Kimi Code
• Start with DeepSeek in Codex
• Choose a model first