Use DeepSeek-V4-Flash in Codex
Choose DeepSeek, Kimi, Qwen, GLM, MiniMax and other available text models from your Codex model picker.
Selected model: DeepSeek-V4-Flash. The copyable prompt below is already customized for it.
Verified October 22, 2026 · Codex CLI 0.162.1 and app-bundled 0.163.0-alpha.10.2 · Requires Node.js 20+, Windows PowerShell 5.1+ and an API key.
1. Copy this prompt into Codex
Let Codex inspect the installer and configure your local client. Enter your API key only at the local hidden prompt. The setup fetches your account's current model list; the public catalog may differ from your account.
Set up ChineseModels in my Windows Codex installation so I can select every available text/chat model and its supported thinking options. API base URL: https://chinesemodels.ai/v1 Model discovery: GET /models with Authorization: Bearer <my key> Chat endpoint: POST /chat/completions Default model: DeepSeek-V4-Flash Guide: https://chinesemodels.ai/documentation/codex Windows adapter: https://chinesemodels.ai/downloads/chinesemodels-codex-windows.zip Do the setup, not just explain it: 1. Detect my installed Codex version and CODEX_HOME. Back up existing configuration and preserve unrelated settings. 2. Download and inspect the adapter. Use an existing local ChineseModels credential if available; otherwise give me the installer's secure local key-entry step. Never print my key or put it in documentation, source code, or a browser URL. 3. Run install-windows.ps1. It prepares a separate profile and a loopback-only Responses-to-Chat-Completions adapter; it does not activate the main app. Store the key using Windows DPAPI. 4. Refresh the authenticated model list. Include all text/chat candidates, preserving exact model IDs. Exclude image and video generators, including any incorrectly tagged as chat. A listed model is not proof of Codex tool compatibility. 5. Configure model-specific thinking choices from capabilities.mjs. Do not show a thinking-off option for always-thinking models or invent effort tiers. Preserve reasoning and complete assistant/tool history across turns. 6. Test text, thinking, streaming, namespaced/custom tools, and a tool-result round trip. Verify model/list using my actual Codex binary and configuration. Report untested models or limitations honestly. 7. Preserve my main OpenAI provider and catalog so Astra and other OpenAI models remain available. Use the separate ChineseModels profile. Do not merge the provider/catalog into my main config: this replaces the normal picker rather than adding a second provider. 8. Give me the launch command: codex --no-daemon --profile chinesemodels --model "DeepSeek-V4-Flash". In that fresh session use /model to select a model and effort. Normal codex sessions retain OpenAI models. Give me the model-refresh command. Keep the full upstream model list separate from the text-model picker. Do not change my sandbox/approval settings or publish anything.
2. Or install from PowerShell
- Download the ZIP above, extract it, and inspect the included scripts. Open PowerShell in that folder.
- Run the command below and enter your ChineseModels API key at the hidden prompt.
- The installer creates a separate ChineseModels profile. Your main Codex provider stays as it was until you activate the profile.
powershell -NoProfile -ExecutionPolicy Bypass -File .\install-windows.ps1
codex --no-daemon --profile chinesemodels --model "DeepSeek-V4-Flash"
The adapter stores your API key with Windows DPAPI. A local authenticated helper runs on 127.0.0.1:18473 and starts on demand. Keep its private settings and encrypted credential out of shared folders and source control.
Keep Astra and the normal OpenAI picker: leave your main config.toml provider and catalog unchanged. Launch ChineseModels with the separate profile command above. This tested setup uses one provider per session; merging its catalog into the main configuration replaces the normal picker. A profile does not add ChineseModels to the default desktop picker.
Download SHA-256 checksum. The archive contains source and installer files, with no customer credentials.
3. Select a model and thinking mode
Close and reopen Codex, start a fresh conversation, then type /model in the terminal. Select the model and one of its listed effort choices. Model IDs are case-sensitive; use the exact ID from your authenticated catalog.
| Family / example | Adapter choices |
|---|---|
| DeepSeek V4 | Off, low, high, max |
| Kimi K3 | Low, high, max; always thinking |
| GLM 5.3 / Kimi K2.7 Code / MiniMax M2.x | Always thinking; no off option |
| Supported Qwen / Hy3 / other toggle families | Off or on; “high” means on |
| MiniMax M3 | Off or adaptive thinking |
| Other models | Model-specific options or provider default |
The installer's capabilities.mjs defines these mappings. They follow model-family documentation; not every model and option has been tested live. Do not interpret a picker entry as certification of tool use. MiniMax adaptive mode can answer simple prompts without reasoning.
4. Refresh, test and troubleshoot
powershell -NoProfile -ExecutionPolicy Bypass -File "$env:USERPROFILE\codex-chinesemodels\refresh-models.ps1"
Restart Codex after refreshing. The model catalog loads at startup. A missing ChineseModels entry often means the session was launched without --profile chinesemodels or predates the refresh. If Astra or other OpenAI models disappear, remove the ChineseModels provider/catalog overrides from the main config using your backup; keep them in the separate profile.
- Authentication errors: check the key, account access and credit balance. Do not put the key in a support screenshot.
- Model not found: fetch /v1/models with your key and use the returned ID. Availability can change.
- Thinking or tool errors: use a supported option and test a text reply, then a tool call and its result. A text reply alone does not verify coding-agent compatibility.
- Restore: use the timestamped configuration backup. Review later edits before restoring so you do not discard unrelated changes.
Our October 22 check discovered 65 text/chat candidates and excluded 28 image/video generators. Your account may return a different list. Codex listed all 65, local routing checks covered 165 model/effort combinations, and representative live requests across seven families succeeded. A DeepSeek tool-result round trip also passed; all 65 models have not been individually certified.
Scope and references
This Windows adapter supplies the Responses interface used by this Codex setup over the gateway's Chat Completions endpoint. It supports text and tool-message translation, with encrypted reasoning replay. It consumes upstream streaming but emits completed output in batches. It does not implement vision input, image/video generation, hosted web search, audio, server-side response history or remote /responses/compact. Its conservative 32,768-token context setting is an adapter limit, not a model specification. Start fresh when switching models with reasoning history.
macOS and Linux users can use the prompt as a checklist, but the included PowerShell/DPAPI installer is Windows-only. A platform-appropriate credential store and launcher are required.