Overview
AstraFlow (星图) is UCloud's enterprise AI development platform. It has two main parts: Modelverse (also called UModelVerse in the docs), a model gateway that lets you call language, image, and video models from many providers with one API key and an OpenAI-compatible interface; and Agent Sandbox, an isolated cloud environment for code execution and agent actions. UCloud's campaign page says the catalog covers 200+ models from 15 providers. Treat that as marketing copy and check the live list in the Modelverse catalog or the console.

The product home is https://astraflow.ucloud.cn/. The Chinese campaign page is https://www.ucloud.cn/site/active/astraflow. Prices are listed at the UModelVerse price page. Docs start at What is UModelVerse and Quick start. The international campaign page is https://www.ucloud-global.com/en/promotion/astraflow. The desktop client is currently branded AstraWork; the download page is https://astraflow-desktop.modelverse.cn/download/.
Do not confuse this product with the open-source project Infini-AI-Lab/astraflow on GitHub. That repository is a reinforcement-learning framework for training language models and agents, not UCloud's cloud platform.
Trade coverage in April 2026 reported that UCloud launched AstraFlow. Features, model IDs, and prices change. The description below is based on the official pages and docs above, checked on 2026-11-13 (Asia/Shanghai). Confirm the live pages before you integrate or pay.
Key uses
- One key for many models: Call DeepSeek, Qwen, GLM, Kimi, MiniMax, Doubao, and other families without opening a separate account and key for every provider.
- Switch models with less refactoring: The platform is OpenAI-compatible. Projects that already use an OpenAI SDK often only need to change the `base_url`, API key, and model ID. Official materials also mention LangChain and LlamaIndex, plus endpoints such as `/v1/chat/completions` and `/v1/models`.
- Language, image, and video in one catalog: The campaign page groups models into LLMs, image generation, and video generation, and names families such as Tongyi Wanxiang, Doubao Seedream, Kolors, Vidu, Kling, Cicada, and MiniMax-H3. UCloud says popular new models can appear within hours; the catalog on the day you call is the source of truth.
- Usage and cost tracking: API calls are billed by tokens or by each model's own unit. Official materials describe usage views at the API key and model level, which helps teams compare quality against cost.
- An isolated runtime for agents: The official site describes Agent Sandbox as a production-style environment for AI agents, with code execution, a graphical desktop, and cloud-side operations, isolated from the host. Use this when the model needs to run code or complete multi-step actions, not only return text.
- A local workspace: The AstraWork download page lists macOS, Windows, Linux, and iPhone. Public materials also mention use with local tools such as OpenCode, Claude Code, and Codex.
Who it is for
- Developers and teams that need to try or switch among several Chinese and other third-party models, especially if the app already speaks an OpenAI-style API.
- People comparing DeepSeek, Qwen, GLM, Kimi, MiniMax, and similar models on quality, latency, and price, without building a new integration for each vendor.
- Products that mix chat with image or video generation and want one entry point first.
- Organizations with residency or compliance needs. UCloud says the platform meets China's Multi-Level Protection Scheme Class III (等保三级). The campaign FAQ also mentions private deployment and VPC access for enterprise customers. The docs provide Chinese and overseas unified endpoints, plus regional endpoints in Ulanqab, Shanghai 2, Los Angeles, Singapore, and Frankfurt.
- Teams that want an isolated place to run agents, execute code, or operate a desktop should evaluate Sandbox in addition to the model API.
If you only want to chat in a browser and do not need an API, DeepSeek or a model maker's own chat app will be lighter. AstraFlow is built as a unified invocation and governance layer. If you care more about a broad international catalog and cross-cloud routing, compare it with this site's OpenRouter page. The category is similar; the company, catalog, and compliance story are not.
How to use it
- Open the AstraFlow site or the campaign page and register a UCloud account. The campaign page says real-name verification is required before you can unlock model access.
- Create an API key in the Modelverse console. Keep the key in your own environment; do not commit it to a public repository.
- Configure three values in your app or tool: the endpoint, the API key, and a model ID. The docs recommend the unified endpoints: `https://api.modelverse.cn` for mainland China and `https://api.umodelverse.ai` for overseas traffic. UCloud says it can fail over a unified hostname to a backup region if needed. If you need a fixed region, use a regional hostname such as `api-cn-wlcb.modelverse.cn` (Ulanqab), `api-cn-sh2.modelverse.cn` (Shanghai 2), `api-us-ca.umodelverse.ai` (Los Angeles), `api-sg.umodelverse.ai` (Singapore), or `api-ge-fra.umodelverse.ai` (Frankfurt).
- Do not guess model IDs. Copy them from the catalog or from `GET /v1/models`. Some families have different IDs per region, and billing follows the ID you actually send.
- If you already use the OpenAI Python SDK, point `base_url` and `api_key` at AstraFlow and start with a `chat.completions` request. The official quick start and authentication guides have the full steps.
- For the desktop app, use the AstraWork download page. Sandbox signup and start flow should be taken from the console on the day you use it.
Pricing
AstraFlow does not publish one flat subscription that covers every model. The campaign FAQ says billing is pay-as-you-go, usually by token, with different prices per model; you pay for what you use, with no minimum spend; settlement can be daily or monthly; and enterprise customers can talk to sales about discounts or committed monthly and annual plans. The site also shows other price styles, including a single rate, context-length tiers, peak and off-peak rates, and per-image or per-duration charges for media models.
The official billing doc adds more modules: fine-tuning is pay-as-you-go by actual training time (hourly, billed by the minute when under an hour); service management is prepaid from compute units and replica count; object storage is pay-as-you-go. Token charges are booked into the hourly bill that contains the request's end time.
On free credits, the billing doc is explicit: after 10:00 on 3 July 2026, new users no longer receive free token credits, and any free credits already granted to existing users were cleared. The doc does not state the time zone; Beijing time is the usual UCloud convention, but use the wording on the page itself. On 2026-11-13, the AstraFlow homepage still invited new users to get "free Modelverse inference credits and Sandbox trial access," which does not fully match the billing doc. Whether any trial or sandbox allowance still exists should be confirmed in the signed-in console and on the price page. Do not treat older figures such as "500,000 free tokens," or the homepage headline, as the current policy.
The SLA states that service availability for the AstraFlow platform solution is no less than 99.00%, effective 1 May 2026. For workloads that cannot accept performance variability or throttling, the SLA suggests a committed-use option. Read the agreement for remedies and exclusions.
This page does not list per-model unit prices, because they change. Check the price page for the model you plan to use before you rely on a cost estimate.
Tips
- Pick the model category from the task. Chat, reasoning, coding, image generation, and video generation are not interchangeable.
- Keep the base URL, key, and model ID configurable so you can change models or regions later without touching business logic.
- When you compare models, freeze the same prompts and look at quality, latency, and the billing unit together.
- Before production, confirm whether the target model is billed per million tokens, by context-length tier, by peak and off-peak hours, or per image or second. Long context and media calls are easy to mis-estimate.
- If you have residency or latency constraints, use the regional model ID and its listed price. Do not assume every call uses the same region or rate.
- Split keys by project or environment so usage is easier to audit and a leak is easier to contain.
Limitations
- The catalog, versions, regions, and prices change. Treat any list on a third-party page as a snapshot.
- You generally need a UCloud account and real-name verification; access may be incomplete without them.
- "200+ models" is the official marketing figure. The models available to a given account, region, or authentication state may be a shorter list.
- OpenAI compatibility does not mean every official OpenAI endpoint or every third-party default will work unchanged. Run a minimal request after you migrate.
- If the homepage, campaign FAQ, and billing doc disagree, follow the console, the price page, and the billing doc.
- This page covers UCloud AstraFlow (星图), not the GitHub project with the same English name.
Frequently asked questions
What is AstraFlow?
AstraFlow (星图) is UCloud's enterprise AI development platform. Its core pieces are the Modelverse model service and Agent Sandbox. For most developers, the usual path is one API key and an OpenAI-compatible API that reaches many models.
Where is the official site?
The product home is https://astraflow.ucloud.cn/. The Chinese campaign page is https://www.ucloud.cn/site/active/astraflow, and the international campaign page is https://www.ucloud-global.com/en/promotion/astraflow. Create keys in the Modelverse area of the UCloud console, not only on a marketing page.
How do I start calling the API?
Register a UCloud account, complete the required verification, create a key in the console, then set `base_url` (usually `https://api.modelverse.cn` in mainland China), your key, and a model ID. See the official quick start for the full sequence.
Which models does AstraFlow support?
The campaign page lists 200+ models from 15 providers, including BAAI, Baidu ERNIE, Bilibili, ByteDance, DeepSeek, EasyLink, HappyHorse, Kuaishou, MiniMax, Moonshot, Qwen, Vidu, Zhipu, Xiaomi, and Cicada, across language, image, and video. On 2026-11-13 the featured list included DeepSeek-V4 models, Qwen3.7, GLM-5.3, Kimi-K3.1-Pro, and MiniMax-H3. The live catalog, console, and `/v1/models` response are authoritative.
Is it compatible with the OpenAI API?
Yes. Official pages say the API is fully compatible with the OpenAI protocol. The usual migration is to point an OpenAI SDK at `https://api.modelverse.cn` or the overseas host `https://api.umodelverse.ai`, then pass your key and a model ID. The docs list `/v1/chat/completions`, `/v1/models`, and `/v1/response` among the compatible endpoints.
How does pricing work, and is there still a free credit grant?
Model APIs are billed as you go, with different unit prices and billing rules per model. Official FAQ text says there is no minimum spend, and settlement can be daily or monthly. See the price page for current rates. On free tokens, the billing doc says that after 10:00 on 3 July 2026, new users no longer receive a free token grant, and existing free grants were cleared. If the homepage still mentions free credits or a trial, confirm what your signed-in account actually shows.
How is AstraFlow different from the UCloud console?
This is a question from the official FAQ. The UCloud console is the general cloud console for compute, storage, and networking. AstraFlow (ModelVerse) is the model-service area inside that console. In AstraFlow you create keys, browse models, and manage usage and cost. Cloud resource management still happens in the main console.
Is AstraFlow the same kind of product as OpenRouter?
Both are developer platforms that aggregate many models behind one API, but they are not the same company and do not offer the same catalog. AstraFlow is operated by UCloud and puts more weight on domestic compliance and regional endpoints. OpenRouter is a separate aggregator. Compare the models you need, the price, data residency, and the account system, not only the "one key, many models" headline.
Is there a client?
Yes. The current official download page brands the desktop app as AstraWork, at https://astraflow-desktop.modelverse.cn/download/, and lists macOS, Windows, Linux, and iPhone. Public materials also mention use with OpenCode, Claude Code, and Codex. Check that page and the docs for what each client can do.
Is this the same as AstraFlow on GitHub?
No. This page is about UCloud's AstraFlow (星图) cloud platform. Infini-AI-Lab/astraflow on GitHub is an open-source reinforcement-learning framework that happens to share the English name.

