GLM-5.3-FlashX API Guide: 200 Tokens/s Inference

Introduction

On September 18, 2026, Zhipu AI officially rolled out GLM-5.3-FlashX, a high-speed inference variant built on GLM-5.3-Flash, the original native multimodal model with 320B total parameters and 18B activated parameters. Through optimized inference infrastructure, the new model achieves a peak output speed of 200 tokens per second. This article dissects the core distinctions between FlashX and its predecessor GLM-5.3-Flash, explains the underlying domestic chip cluster architecture that enables this speed boost, and provides practical API integration steps, scenario analysis and benchmark data for developers and enterprise teams.

Core Facts Overview

The key performance and commercial metrics of GLM-5.3-FlashX are sourced from Zhipu AI official releases and market news published on September 18, 2026:

1. Peak inference speed reaches 200 tokens/s, representing a 5x speed increase compared with GLM-5.3-Flash.
2. The pricing of FlashX is set at 2.5 times the original Flash version, forming a three-way tradeoff triangle of intelligence, cost and latency.
3. The base GLM-5.3-Flash model has 320B total parameters and 18B activated parameters. Its Artificial Analysis intelligence index hits 57 points, matching the score of Claude Opus 4.8. This benchmark data was released by Zhipu AI on August 26, 2026.
4. Before the official launch, GLM-5.3-Flash was tested under the alias Ox Alpha in OpenCode and OpenRouter. It quickly topped the usage volume ranking on both platforms. On OpenRouter alone, it handled more than 2 trillion tokens of requests within a 6-day testing cycle.
5. Inference workloads run on a cluster of 100,000 domestic AI chips. Powered by an EPD separated pipeline architecture, end-to-end service performance is lifted by 3 times versus the baseline deployment.

GLM-5.3-FlashX targets enterprises and developers who demand ultra-fast inference. Its API service went live simultaneously with the model release, with the designated Model Key GLM-5.3-FlashX. The original GLM-5.3-Flash gained broad recognition from the global developer community under the Ox Alpha alias, with API traffic rising continuously. To meet surging real-time request demands, Zhipu AI expanded its inference infrastructure on top of the existing 100,000 domestic chip fleet. FlashX preserves the same intelligence capability of GLM-5.3-Flash while drastically cutting generation latency, creating a competitive balance among model intelligence, pricing and response speed.

GLM-5.3-FlashX vs GLM-5.3-Flash: Core Differences

GLM-5.3-FlashX is a variable-speed optimized iteration. It retains identical base model intelligence while maximizing token generation throughput. The base GLM-5.3-Flash is the first native multimodal model in the GLM-5 family, with 320B total parameters and 18B activated parameters. Three core dimensions separate FlashX and Flash:

This speed upgrade is particularly valuable for frequent interactive workflows. A typical Agent loop takes around 2–3 seconds per turn. After 10 rounds of iteration, the accumulated waiting time can reach half a minute. FlashX shrinks the per-turn generation delay significantly and accelerates the convergence of multi-step agent tasks.

Model Foundation: Why FlashX Can Boost Speed Without Sacrificing Intelligence

The strong baseline capability of GLM-5.3-Flash is the prerequisite for FlashX’s speed upgrade. The model balances large-scale parameters, sparse activation and low inference overhead:

1. Parameter Design: Total parameters stand at 320B, close to the 355B of GLM-4.5. But activated parameters drop from 32B to 18B, and layer count reduces from 92 to 45, cutting nearly half of computation overhead.
2. Pre-training Corpus: Trained on a 30T token multimodal prediction dataset. It achieves robust capability with less computation, according to Zhipu AI’s technical blog posted on August 26.
3. Architectural Innovation: It is an open-source frontier model combining sparse attention and linear attention. The built-in IndexPool compresses 1M context retrieval indexes, reducing attention computation volume by 3.01 times and cutting KV memory consumption by 4.44 times compared with GLM-5.3.
4. Pricing Benchmark: GLM-5.3-Flash costs 1/10 of the full GLM-5.3 model, with limited-time discounts down to 1/20. Its price is only 1/40 of Claude Opus 4.8, while delivering equivalent intelligence scores.
5. Real-world Load Validation: In large-scale Ox Alpha trials on OpenCode and OpenRouter, it became the most popular model during the testing window. OpenRouter recorded over 2 trillion tokens processed within six days.

How 200 Tokens/s Is Achieved: Domestic Chips and Infrastructural Optimization

FlashX’s speed gain is not from overclocking model weights, but systematic engineering optimization. The 100,000-node domestic chip cluster supplies the raw inference compute, and refined infrastructure tuning unlocks the peak throughput.

• Domestic Chip Cluster: Chips are interconnected via high-bandwidth self-developed network links. This deployment delivers frontier model services on domestic silicon at massive traffic scale for the first time.
• Specialized Inference Engine: Built on top of SGLang, the dedicated inference engine leverages intra-node tensor parallelism, ReplaySSM, W8A quantization, and mixed-precision storage including INT8, FP8 and BF16 to boost layer execution efficiency.
• EPD Separation Architecture: Production-grade Encode–Prefill–Decode decoupling splits multimodal encoding, context prefill and token-by-token decoding into independently schedulable, elastic scaling workload pools.
• Measured Outcome: Compared with the original hardware baseline, end-to-end service capacity rises by 3 times. Per-token hardware cost reaches parity with mainstream NVIDIA GPU deployments.

Developers can connect to the model through the standard OpenAI-compatible SDK. Compatible gateway endpoints allow switching between multiple large models without rewriting existing business code. When managing multiple model endpoints for mixed workloads, 4sapi, an API gateway, simplifies unified routing and traffic governance.

GLM-5.3-FlashX Integration: Three Steps to Get Started

The API of FlashX maintains full compatibility with GLM-5.3-Flash. Users only need to modify the model key parameter, while chat completion interfaces and multimodal field definitions remain unchanged.

1. API Endpoint Access: Official documentation is hosted at https://docs.bigmodel.cn/api-reference/model-api/chat-completion. The base_url is https://open.bigmodel.cn/api/paas/v4/.
2. Web Trial: Developers can test the model visually at https://www.bigmodel.cn/trialcenter/modeltrial/visual?modelCode=glm-5.3-flashx.
3. Model Parameter: In API requests, fill the model field with glm-5.3-flashx.

from openai import OpenAI

client = OpenAI(
    api_key="<your-key>",
    base_url="https://open.bigmodel.cn/api/paas/v4/"
)

resp = client.chat.completions.create(
    model="glm-5.3-flashx",
    messages=[],
    stream=True,
    max_tokens=1024
)

Three critical notes for integration:

• Streaming output best demonstrates the 200 tokens/s speed advantage. Always enable stream:true.
• Set max_tokens ≥1024 to avoid output truncation.
• Native multimodal support accepts image_url, video_url and file input fields directly.

Suitable and Unsuitable Workloads for GLM-5.3-FlashX

Recommended Scenarios for FlashX

FlashX is optimized for latency-sensitive and frequently invoked workloads:

1. AI Programming and ZCode Integration: Code completion, refactoring and long-file rewriting. Streaming responses drastically reduce waiting time, and the model has been integrated into coding platforms such as ZCode.
2. Agent Multi-step Loops: Every tool call round is shortened to sub-second latency, accelerating overall task convergence.
3. Real-time Assistants and Experience Centers: Typing echo delay is minimized, delivering near-human response experience.
4. 1M Long Context Tasks: Mixed attention architecture preserves precise long-context reasoning with low KV cache overhead.

Scenarios Where Standard GLM-5.3-Flash Works Better

For offline batch processing, workloads insensitive to token latency, and heavy-volume requests with strict budget constraints, GLM-5.3-Flash remains more cost-effective. FlashX is priced at 2.5 times Flash, so teams need to calculate total cost based on target QPS and token volume.

Frequently Asked Questions

Q: Does GLM-5.3-FlashX have the same intelligence as GLM-5.3-Flash?

They share the identical base model intelligence. FlashX only accelerates inference through infrastructure optimization. Metrics including 320B total parameters, 18B activated parameters, AA score of 57 and 30T pre-training corpus are consistent. The only differences lie in generation speed and billing price. Selection depends on latency budget.

Q: What is the relationship between Ox Alpha and FlashX?

Ox Alpha was the anonymous test alias of GLM-5.3-Flash during the OpenCode and OpenRouter beta period, and it broke usage records on both platforms. FlashX is the official speed-enhanced release of Flash. Both run on the same domestic chip cluster infrastructure.

Q: How stable is inference on domestic chips?

According to Zhipu AI’s August disclosure, the EPD separated architecture lifts end-to-end throughput by three times. Its per-token cost matches mainstream GPU deployments. The system has supported a 2 trillion-token workload over six days on OpenRouter and meets large-scale production requirements.

Q: Do existing Flash API users need to rewrite their code?

No code rewrite is required. Change the model parameter value to glm-5.3-flashx. Chat dialogue, tool invocation and multimodal fields are fully compatible. It is recommended to turn on streaming mode to observe the difference in generation speed.

Q: How can students and individual developers test at low cost?

Developers can use the web visual trial platform, or join the GLM Coding Plan, which releases 10,000 trial tokens daily. Validate model performance first before enabling paid API traffic.

Conclusion

GLM-5.3-FlashX delivers up to 200 tokens/s generation speed, a 5x throughput increase at a 2.5x pricing premium. It brings the full intelligence capability of GLM-5.3-Flash into real-time interactive scenarios. Built upon a 100,000-node domestic chip fleet and EPD separated pipeline design, the base model achieves an AA score of 57, on par with Claude Opus 4.8. All information in this article is sourced from public materials released by Zhipu AI on September 18, 2026. Developers should cross-verify pricing and rate limits from Zhipu’s official documentation before production deployment.

International access: https://4sapi.com
Domestic access: https://4sapi.cn