Blog / Relays & gateways

Alibaba Cloud Model Studio Adds GLM-5.3-FlashX and step-5-preview

2026-10-19cdnping editors4 min read

On September 21, 2026, the "Newly released models" page of Alibaba Cloud Model Studio (known in China as Bailian) recorded three third-party models going live on the same day: ZHIPU/GLM-5.3-FlashX, stepfun/step-5-preview, and vanchin/deepseek-v4.1-flash. If you are looking up how to call GLM-5.3-FlashX on Model Studio, start with one point: the same model can have a different name, rate limit, and latency on the official API than on Model Studio, so check each model code before you call it. Everything below is current as of October 2026. Sources are the Model Studio newly released models page and the Zhipu Open Platform changelog.

#What went live on September 21?

According to the Model Studio page (North China 2, Beijing region):

A few notes. The step-5-preview page does not give an output length, and I am not filling in a number for it. The DeepSeek row above is the entry with the vanchin/ prefix. It is a different entry from deepseek-v4.1-flash, which was listed on September 13 and is described on the page as a 552B-parameter MoE. Earlier listings include kimi-k3 (August 19) and glm-5.3 (the page records August 28 for both the China site and the Singapore site).

Zhipu's own changelog lists GLM-5.3 (2026-08-19) and GLM-5.3-Flash (2026-08-26). I did not find a GLM-5.3-FlashX entry on that page, so its speed and specs rest on the Model Studio page alone. Check the model detail page to confirm.

#What should you watch for with the same model on the official API vs. Model Studio?

Don't assume that the same name means the same behavior. These models go live quickly, so each platform's onboarding date, rate limits, and default parameters are likely to differ. Before switching endpoints, check these items one by one:

1. Model code: Model Studio uses a vendor prefix (such as ZHIPU/, stepfun/, vanchin/), while official APIs usually don't. Don't copy a name across.
2. Context and output limits: use the model detail page of each platform, not the numbers from another platform.
3. Rate limits and concurrency: each platform counts them separately. See the console for the actual values.
4. Billing: official and third-party platforms price separately. Use each platform's pricing page.
5. Input types: whether images, video, and files are supported depends on each platform's docs.

#How do you pick an endpoint for your own network?

When the model is the same, connection quality to the endpoint often decides how the service feels. You can measure it like this:

1. Open the cdnping home page and test the API endpoints of DeepSeek, Zhipu, Kimi (Moonshot AI), Alibaba Cloud Model Studio, Volcengine Ark, and SiliconFlow side by side.
2. Look at two numbers: "first connect" (includes DNS and the TLS handshake) and "round trip" (the median of 3 requests after the connection is reused). Long-lived connections and streaming calls depend more on round trip. Scripts that open a new connection for every request depend more on first connect.
3. Test several times and at different times of day. Don't draw conclusions from a single result. For how we measure, see the methodology. For where latency comes from, see AI API latency from China.
4. To check whether an endpoint is healthy right now, see /en/status/dashscope, /en/status/deepseek, and /en/status/zhipu.

To compare the same model across inference providers, use /en/models. It is built on the public OpenRouter API. For the latest models from each vendor, it shows per-provider availability, latency, throughput, and quantization (such as fp8 and fp4), and it records listings, delistings, price changes, and quantization changes. Note that this is OpenRouter data and does not describe Model Studio.

#How do you write a unified interface?

For the exact address and authentication method of Model Studio's compatible API, follow the official docs. The general approach is to keep the base URL, API key, and model name in configuration instead of hard-coding them:

export LLM_BASE_URL="<see the official docs>"
export LLM_API_KEY="<your key>"
export LLM_MODEL="ZHIPU/GLM-5.3-FlashX"

Then switching endpoints only means changing environment variables. Official model names keep changing too. For how to handle that, see old model name migration and the DeepSeek V4.1-Flash-1017 update. For DeepSeek pricing details, see peak and off-peak pricing.

If you use tools such as Claude Code or Codex, see custom API configuration for setting a custom base URL. Model Studio also released other new models around the same time. See the Model Studio September-October new model roundup.

#What if you don't want to integrate each vendor separately?

Opening accounts with each vendor and comparing their limits takes real effort. AGI Plan is a commercial partner of cdnping. It offers a monthly subscription, billed in RMB, to a multi-provider aggregation gateway, with capability groups such as "Chinese flagship models" that suit long Chinese-language text and batch jobs. It is not an officially authorized channel for any model. Available models and regions are listed at https://agiplan.dev/models. Whether it fits you is something to decide from your own tests.

#Sources

• Alibaba Cloud Model Studio: Newly released models (accessed October 2026)
• Zhipu Open Platform: Changelog