GLM-5.3-Flash: Vision, API Pricing and Local Setup
GLM-5.3-Flash is Z.ai’s open-weight multimodal model for visual coding, document work and tool-using assistants. Choose it when screenshots or images belong in the task, or when you want to test a lower-cost GLM route. It uses a different 320B-total, 18B-active architecture from GLM-5.3 and produces text.
Content reviewed:
Use this model
API docs and console
- Read the GLM-5.3-Flash API guide
Provider account and usage billing apply.
Official model repository
- Official weights and framework links
Model card, downloads and license conditions.
Choose a first task
- Turn a screenshot and a written brief into a frontend change, then inspect the rendered result.
- Read a document or visual reference and produce a structured text answer.
- Run a small coding-agent task with an explicit test and output budget.
When should you choose GLM-5.3-Flash?
Start with Flash when the task needs visual evidence: a UI screenshot, a document page or feedback from a browser. Its output is text; creating a finished slide deck or operating an interface also needs an application that executes tools and handles files.
For text-only repository work, compare Flash with GLM-5.3 using the same prompt, tools, reasoning effort and acceptance test. A lower token price is useful only if the model finishes the task with an acceptable number of attempts.
How much does GLM-5.3-Flash cost?
Z.ai lists $0.15 per million input tokens, $0.03 for cached input and $0.50 per million output tokens. For an uncached request with 20,000 input and 3,000 output tokens, the calculation is 0.02 × $0.15 + 0.003 × $0.50 = $0.0045. A thousand such requests cost $4.50 before tools and other charges.
The Flash name does not mean free API access. The price page separately lists free older Flash models. Cache-storage pricing is a limited-time free offer, separate from the cached-input rate. Z.ai lists its built-in web search at $0.01 per use; external tools can add their own charges.
| Billing period | Uncached input | Output | Example total |
|---|---|---|---|
| Standard | $0.15 | $0.50 | $0.20 |
Example: 1M uncached input tokens + 100K output tokens.
Official pricing · Source checked: · Standard
Flash vs GLM-5.3: compare the same workload
For a screenshot-driven change, GLM-5.3-Flash can read the reference directly; GLM-5.3 needs a text description or a separate visual-processing step. Include that extra step when comparing the complete workflow, rather than pricing only the final coding request.
The table uses identical uncached token counts to show the billing difference. A visual task can have a different input-token count from a text-only task. Record the actual usage for the image request and the follow-up edits, then check the rendered interface against the reference.
| Choice | GLM-5.3-Flash | GLM-5.3 |
|---|---|---|
| Input → output | Text, image, video and file inputs → text | Text → text |
| Context / maximum output | 1M / 128K tokens | 1M / 128K tokens |
| Input / output per 1M tokens | $0.15 / $0.50 | $1.40 / $4.40 |
| Uncached 20K input + 3K output | $0.0045 | $0.0412 |
| Weight license | MIT | GLM-5.3 License |
Send a first GLM-5.3-Flash image request
Send GLM-5.3-Flash a screenshot and a question as image_url and text blocks in messages[].content. Replace the example URL with an accessible screenshot of your interface. Z.ai recommends an image URL; a Base64 data URL is another supported option. Add one image_url block per image for a multi-image request.
Use a server-side Z.ai API key for the pay-as-you-go endpoint below. Begin with a short visual-inspection task and low reasoning effort; expand the output budget when moving on to implementation. GLM Coding Plan access uses a separate subscription endpoint and allowance.
curl https://api.z.ai/api/paas/v4/chat/completions \
-H "Authorization: Bearer $ZAI_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "glm-5.3-flash",
"messages": [
{
"role": "user",
"content": [
{
"type": "image_url",
"image_url": {
"url": "https://example.com/your-interface-screenshot.png"
}
},
{
"type": "text",
"text": "Describe the navigation, spacing and typography in this screenshot. List three details to preserve when implementing it."
}
]
}
],
"thinking": {
"type": "enabled"
},
"reasoning_effort": "low",
"temperature": 1,
"top_p": 0.95,
"max_tokens": 4096
}'How did Kong evaluate GLM-5.3-Flash?
Kong evaluated GLM-5.3-Flash alongside GLM-5.3 for its code-review workflow. See the GLM-5.3 guide for the evaluation conditions and the separate effect of routing reviews by complexity.
Set GLM-5.3-Flash reasoning and output budgets
GLM-5.3-Flash keeps reasoning enabled and accepts low, high or max effort, defaulting to max. For a visual coding loop, first ask for an inspection of the screenshot, then request the implementation and provide a new screenshot of the rendered result. This makes the next correction specific to what the model can see.
Reserve enough output for both reasoning and the requested code or explanation. If the response ends before the implementation is complete, inspect the completion limit and finish reason before requesting another take. Keep the screenshot set and effort setting fixed when comparing changes.
FlashX and FlashX-Legal are separate API IDs with their own prices. Select glm-5.3-flash for the standard Flash rate; FlashX’s advertised throughput does not describe this endpoint.
What does GLM-5.3-Flash local serving require?
The official Flash checkpoint is FP8. Its 18B active parameters describe the work selected per token, not the full set of weights that must be available to the runtime. Reserve memory for the whole checkpoint, its hybrid attention state, visual inputs and concurrency.
The vLLM recipe specifies a model-aware nightly image containing native GLM5Next support and its Rust frontend. Follow the recipe for your GPU platform rather than assuming a generic stable image can serve the checkpoint. Begin with short context and one request before raising concurrency.
SGLang’s SM120 tracking issue records configuration-specific fixes and qualification work for Flash. Check the exact runtime and kernel revisions for your machine. A successful weight load is only the first check; test image inputs, repeated tool calls and sustained requests before sizing a service.
Which license applies to the weights?
GLM-5.3-Flash weights use the MIT License. Preserve the required copyright and permission notice when redistributing them. Hosted API usage also follows the service’s terms and billing conditions.
GLM-5.3 uses a separate custom license. Keep the two licenses separate when deciding which checkpoint to redistribute, fine-tune or host for others.
Before you use it
- A visual answer is text; creating files, running the application and inspecting a new screenshot require a client that supplies those tools.
- FlashX and FlashX-Legal use separate model IDs and billing. A client switch to either route changes the price from the standard Flash quote.
How to get started
Send one screenshot with a specific inspection question, then implement and render the change. Compare the new screenshot with the reference and count all image requests, code responses and retries in the cost of the accepted result.
Sources
Sources (8)
- Kong code-review workflow and internal evaluation ↗
Source checked:
- Official model guide ↗
Source checked:
- Z.ai API pricing ↗
Source checked:
- Chat Completions request parameters ↗
Source checked:
- Official model card and deployment links ↗
Source checked:
- Weight license ↗
Source checked:
- Runtime deployment recipe ↗
Source checked:
- Configuration-specific community runtime report ↗
Source checked:
Calculators
Using this calculator for this model
Estimate a text model’s ongoing cost by choosing where it runs and how many hours it is active each month. Usage levels are 120, 360 or 720 hours per month; choose the closest to your planned schedule.
For open weights, select the precision and deployment that match the artifact and runtime you plan to use. The hardware result is a capacity estimate; context length, runtime and model-specific components can change actual memory use.
How to fill in the calculator
- Precision
- Select the weight format you will actually run. GGUF Q4 is a broad capacity estimate for a 4-bit GGUF variant; exact file size depends on the quantization and included components.
- Deployment
- Local estimates use hardware ownership, electricity and operations; cloud estimates use hourly rental pricing. Compare the path you are considering.
- Usage level
- Choose 120, 360 or 720 active hours per month, based on your planned schedule. For bursty work, use the nearest level as a comparison.
- Electricity rate
- Use your billed price per kWh. This changes local electricity cost only.
- Wall power
- Leave blank to use the GPU TDP reference. For multiple GPUs or a full system, enter measured whole-machine watts.
- Salary and FTE
- Enter a monthly salary range and the share of one full-time role spent operating the service; use zero if you allocate no labor.
- PUE and system cost
- Keep defaults if unknown. PUE scales local electricity; the system cost multiplier scales upfront hardware cost.
Use the calculator for a deployment budget; enter your expected usage and measured wall power.