GLM-5.3-Flash

glm-5.3-flash

GLM-5.3-Flash is the first native multimodal model in the GLM-5 series, achieving intelligence stronger than GLM-5.2 with an ultra-low-cost architecture.

Context window1.0M

Provider智谱AI

Released2026/09/01

Playground

Pricing

The same model is available through multiple service channels — choose based on latency, reliability and cost.

Prices in $ / 1M tokens

To pick a channel, add a provider field to the request body, for example "provider": { "channel": "direct" }. Valid values are direct / stable / economical; omit it to use the default channel.

Direct

Direct upstream connection — best when you need native behavior and the full context window.

Input	Output	Cache read
0.15/M	0.50/M	0.03/M

Overview

Input

Text Image

Output

Text

GLM-5.3-Flash API: Hybrid-Attention Multimodal Model That Can See Images and Get Things Done

GLM-5.3-Flash is a model released by Z.ai on August 26, 2026, and open-sourced under the MIT license. It is the first natively multimodal member of the GLM-5 series. With 320 billion total parameters and 18 billion activated per token, it supports image input and a 1 million token context.

What makes it stand out is not its parameter count but its architecture: a hybrid design of linear attention and sparse attention, with layer depth reduced from 92 layers in GLM-4.5 to 45 layers, and the introduction of mHC (manifold-constrained hyperconnectivity). Compared with the flagship GLM-5.3, attention computation is reduced by approximately 3.01x and KV cache by approximately 4.44x—which translates directly into a price tier roughly one-tenth that of GLM-5.3.

On the Artificial Analysis composite intelligence index, GLM-5.3-Flash scored 57, entering the global frontier model tier; it ranks first in its cohort on both the Toolathlon Verified (78.4) and AutomationBench (48.8) agent benchmarks.

SeaWhale AI offers GLM-5.3-Flash through an OpenAI-compatible API, supporting image input, tool calling, reasoning_effort reasoning-intensity control, and streaming output.

Get API Key · Model ID: GLM-5.3-Flash

Why Choose GLM-5.3-Flash

Natively multimodal — The first model in the GLM-5 series to support image input, accepting multiple images at once
AA Intelligence Index score of 57 — Enters the frontier model tier with 18 billion activated parameters
Hybrid attention architecture — Reduces attention computation by approximately 3.01x and KV cache by approximately 4.44x
Priced at about 1/10 of GLM-5.3 — Frontier capabilities at a lightweight cost
MIT open-source license — Fully open weights, free for commercial use and private deployment
1 million token context — Up to 128K token output

Core Capabilities

01 Visual Programming and "Watching Itself Work"

Natively multimodal, the model can read design mockups, screenshots, and error screens directly and then write code—without first requiring another model to convert images into text descriptions. When doing front-end implementation, UI debugging, and automated workflows, the model can see what the interface it produced looks like and then decide what to change next.

Generate front-end code directly from design mockups / screenshots
Visually locate and fix interface defects
Mixed text-image multi-round iterative development

02 Agents and Tool Calling

With Toolathlon Verified 78.4 (first in its cohort) and AutomationBench 48.8 (first in its cohort), it ranks among the top tier in multi-tool orchestration and automated workflows for this parameter class. It scored 1773 on GDPval-AA v2, ranking 2nd on real professional tasks.

Multi-tool orchestration and long-chain automation
Knowledge work on real professional tasks
Mixed text-image agent workflows

03 Efficiency Advantages of Hybrid Attention

The hybrid architecture of linear attention + sparse attention, combined with the shallow 45-layer design and mHC, reduces the two most expensive overheads in long-context scenarios: attention computation to about 1/3 and KV cache to about 1/4.4. This is the direct reason it can price at one-tenth of the flagship, and it also makes long-context inference latency significantly lower.

Low-latency responses under long context
Controllable cost in high-concurrency scenarios
VRAM-friendly for private deployment

04 1-Million-Token Multimodal Context

A 1-million-token context that accepts both text and images, with a maximum output of 128K tokens. An entire illustrated technical document, a full set of design mockups, and the corresponding code repository can all fit at once.

Long-document understanding with mixed text and images
Combined analysis of full repository + UI screenshots
Full retention of long conversation history

Best Use Cases

Scenario	Description
Visual programming	Convert design mockups and screenshots directly into front-end code and UI debugging
High-concurrency agents	Toolathlon and AutomationBench first in cohort, with controllable cost
Mixed image-text document processing	Long-context understanding of illustrated technical documents and reports
Cost-sensitive production deployment	Price tier roughly one-tenth that of GLM-5.3
Private deployment	MIT license + 18 billion activated parameters, friendly VRAM requirements
High-volume processing	Low-latency long context, suitable for pipeline tasks

GLM-5.3-Flash vs. GLM-5.3 and GLM-5.2

Capability	GLM-5.3-Flash	GLM-5.3	GLM-5.2
Model ID	`GLM-5.3-Flash`	`GLM-5.3`	`GLM-5.2`
Total / activated parameters	320B / 18B	753B / ~40B	Same base as GLM-5.3
Model layers	45	—	—
Input modalities	Text + image	Text	Text
Context window	1 million tokens	1 million tokens	1 million tokens
Max output	128K tokens	128K tokens	128K tokens
AA Intelligence Index	57	Higher	Lower than GLM-5.3
Terminal-Bench 2.1	84.3	88.2	Slightly lower
Weights license	MIT	GLM-5.3 License	MIT
Reasoning effort	low / high / max	low / high / max	high / xhigh
Positioning	Lightweight multimodal, ~1/10 price	Flagship: strongest coding and safety	Previous flagship

Specific pricing is subject to the real-time price card at the top of the page.

FAQ

When was GLM-5.3-Flash released?

It officially launched on August 26, 2026, with weights open-sourced at the same time under the MIT license.

How does it relate to GLM-5.3?

It is not a pruned version of the same base, but an independent lightweight multimodal line: 320B total parameters and 18B activated, using a linear + sparse hybrid attention architecture for efficiency optimization, at roughly one-tenth the price of GLM-5.3. GLM-5.3 is a text-only flagship with stronger coding and safety capabilities; Flash adds image input and has advantages in cost and latency.

What is the difference between "natively multimodal" and an externally attached vision model?

Natively multimodal means the model is trained on mixed image-text data during the pretraining stage (about 30 trillion multimodal tokens) so it understands images directly, rather than having another model convert images into text first. That layer of information loss is eliminated, and tasks like visual programming and UI debugging benefit the most.

How much does hybrid attention specifically save?

Compared with GLM-5.3, attention computation is reduced by approximately 3.01x and KV cache by approximately 4.44x. Since these two are the main overheads in long-context scenarios, the reductions in latency and cost are most pronounced with long inputs.

Is it smart enough?

It scored 57 on the AA composite intelligence index, already within the global frontier model tier. Toolathlon Verified 78.4 and AutomationBench 48.8 are both first in their cohort, and GDPval-AA v2 scores 1773, ranking 2nd. In pure coding strength, it falls short of GLM-5.3 (Terminal-Bench 2.1: 84.3 vs. 88.2).

Can the thinking mode be turned off?

No. The thinking mode is always on and can be adjusted among three levels—low/high/max—via reasoning_effort. The official recommendation is max.

Can it be deployed privately?

Yes. It uses the MIT license, weights are approximately 328GB (FP8), and it supports mainstream inference frameworks such as vLLM, SGLang, Transformers, and KTransformers.

Why Choose SeaWhale AI for the GLM-5.3-Flash API

Call directly without deployment — No need to prepare 328GB of weights or a GPU cluster
OpenAI-compatible API — Image input uses the standard multimodal format; change two lines in existing code to integrate
Compare side-by-side with closed-source models — Under the same account, directly compare performance and cost with Claude, GPT, and Gemini
RMB settlement — Domestic models are billed in RMB for clear accounting

API

API integration

Model IDUse this value as the model in inference requests

glm-5.3-flash

API KeyBearer token used to authenticate inference requests

Manage API keys

Base URLOpenAI compatible · /chat/completions

https://api.haijingai.com/v2/

provider OptionalSelects a service channel; omit it and the system picks the default

"provider": { "channel": "direct" }

glm-5.3-flash usage examples

SeaWhale AI is compatible with the OpenAI API protocol, so you can call it with the OpenAI SDK or plain HTTP requests. Streaming is enabled by default.

cURLPythonNode.js

curl https://api.haijingai.com/v2/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer <API_KEY>" \
  -d '{
    "model": "glm-5.3-flash",
    "messages": [
      {"role": "system", "content": "You are a helpful assistant."},
      {"role": "user", "content": "Hello!"}
    ],
    "provider": { "channel": "direct" },
    "stream": true
  }'
# provider is optional — remove this line to use the default channel

from openai import OpenAI

client = OpenAI(
    base_url="https://api.haijingai.com/v2",
    api_key="<API_KEY>",
)

stream = client.chat.completions.create(
    model="glm-5.3-flash",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "Hello!"},
    ],
    stream=True,
    # Optional: pick a service channel; omit to use the default
    extra_body={"provider": {"channel": "direct"}},
)

for chunk in stream:
    if chunk.choices[0].delta.content:
        print(chunk.choices[0].delta.content, end="", flush=True)

import OpenAI from 'openai'

const client = new OpenAI({
  baseURL: 'https://api.haijingai.com/v2',
  apiKey: '<API_KEY>',
})

const stream = await client.chat.completions.create({
  model: 'glm-5.3-flash',
  messages: [
    { role: 'system', content: 'You are a helpful assistant.' },
    { role: 'user', content: 'Hello!' },
  ],
  stream: true,
  // Optional: pick a service channel; omit to use the default
  // @ts-expect-error provider is a SeaWhale AI extension, not in the OpenAI SDK types
  provider: { channel: 'direct' },
})

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? '')
}