Zhipu's GLM-5.3-Flash (model ID glm-5.3-flash) is now live on HopBase, served over the OpenAI-compatible POST /v1/chat/completions endpoint, so any OpenAI SDK works by swapping the model name. It is a 320B-parameter Mixture-of-Experts model that activates 18B parameters per token, and the first natively multimodal member of the GLM-5 family: it accepts text, images, video and files as input and returns text. The context window is 1M tokens with up to 128K output tokens in a single response, and the reasoning trace comes back in a separate reasoning_content field. HopBase bills it against Zhipu's official list price — CNY 0.80 per million input tokens, CNY 0.23 per million cached input tokens and CNY 2.80 per million output tokens, one tenth of GLM-5.3's official pricing. Keys bound to the GLM 5.3 Flash plan group can call it today.
How to call it
- One OpenAI-compatible endpoint:
POST /v1/chat/completionswith base URLhttps://api.hop-base.com/v1. No SDK changes. - Use the full model ID:
glm-5.3-flash. Shortened names do not resolve; read the authoritative list fromGET /v1/models. - Group binding: the model sits in the GLM 5.3 Flash plan group. Only keys bound to that group can reach it; GLM-5.3 keys are not shared.
- Reasoning in its own field: the thinking trace is returned in
reasoning_contentand the answer incontent, so your client does not have to split them apart.
Pricing: at the official list price
| Per million tokens | GLM-5.3-Flash |
|---|---|
| Input | CNY 0.80 |
| Cached input read | CNY 0.23 |
| Output | CNY 2.80 |
These are Zhipu's published list prices; the official pricing is one tenth of GLM-5.3. Charges are list price times actual usage, itemised per request in the console, and your effective unit price depends on your plan group. Cached reads cost less than a third of the input price, so agent loops and long conversations that re-read the same prefix save the most. For enterprise volume, contact sales from the pricing page.
Specs and what we verified
- Context and output: 1M context window, up to 128K output tokens per response.
- Native multimodal input: text, images, video and files in; text out.
- Architecture: hybrid sparse and linear attention. Per Zhipu's own figures it cuts attention compute by 3.01x and KV cache by 4.44x versus GLM-5.3, which is what makes the one-tenth price possible.
- Verified on our production path: non-streaming, streaming (usage arrives with
stream_options.include_usage), function calling (tool_callscomes back intact) and thereasoning_contentthinking trace.
Parameter guidance
- Zhipu recommends
temperature: 1,top_p: 0.95andreasoning_effort: "max". Do not carry over the low-temperature defaults you use for other models. - Thinking mode is always on: there is no switch to disable it, so control response length with your prompt and
max_tokens. - For streaming, enable both
stream: trueandtool_stream: true; addstream_options: {"include_usage": true}if you need usage in the stream.
Three-minute integration
curl https://api.hop-base.com/v1/chat/completions \
-H "Authorization: Bearer sk-your-key" \
-H "Content-Type: application/json" \
-d '{
"model": "glm-5.3-flash",
"messages": [{"role": "user", "content": "Introduce yourself in three sentences"}],
"temperature": 1,
"top_p": 0.95,
"stream": true,
"stream_options": {"include_usage": true}
}'
Full integration notes, tool calling and multimodal examples are in the GLM-5.3-Flash integration guide. If you already have a key, bind it to the GLM 5.3 Flash group and start calling.
FAQ
How do I call GLM-5.3-Flash?
Send an OpenAI-compatible request to POST https://api.hop-base.com/v1/chat/completions with model ID glm-5.3-flash. Any OpenAI SDK works by swapping the model name; authentication is unchanged. Your key has to be bound to the GLM 5.3 Flash plan group, and GET /v1/models is the authoritative list of what that key can reach.
How much does GLM-5.3-Flash cost?
It is served at Zhipu's official list price: CNY 0.80 per million input tokens, CNY 0.23 per million cached input reads and CNY 2.80 per million output tokens — one tenth of GLM-5.3's official pricing. Charges are list price times actual usage, itemised per request in the console, and the effective unit price depends on your plan group. Enterprise volume pricing is available from the pricing page.
Does it support multimodal input and tool calling?
Yes. It is the first natively multimodal member of the GLM-5 family, accepting text, images, video and files as input and returning text. Function calling (tools / tool_calls) is verified on our production path; for streaming, enable tool_stream: true, and the reasoning trace is returned in reasoning_content.
GLM-5.3-Flash or GLM-5.3?
Both offer a 1M context window. Flash is officially priced at one tenth of GLM-5.3, so high-volume workloads — everyday chat, agent loops, long-context batch processing — should start with Flash. Existing workloads tuned against GLM-5.3 can stay on it. They are separate plan groups and keys are not shared, so you can integrate both and shift traffic on measured results.
What is the maximum output length?
Up to 128K output tokens per response, within a 1M context window. Use streaming for long outputs so edge timeouts do not cut the connection, and add stream_options: {"include_usage": true} if you need usage statistics in the stream.