News · Medium impactBack to News

Zhipu AI Launches GLM-5.3-FlashX: 200 Tokens/s High-Speed Inference for High-Concurrency Coding Agents

Zhipu AI officially launched the GLM-5.3-FlashX API on September 18, 2026. Built on a hybrid attention architecture with 320B total and 18B active parameters, the model maintains a 1M context window while accelerating inference up to 200 tokens/s, priced at 2.0 RMB input and 7.0 RMB output per million tokens.

High-Speed Inference Architecture and Version Iteration Background

Zhipu AI officially released the GLM-5.3-FlashX high-performance inference model on September 18, 2026. Previously, the base GLM-5.3-Flash had undergone blind testing under the anonymous codename Ox (Ox Alpha) on OpenCode and OpenRouter, earning widespread praise for its solid coding capability; the newly launched GLM-5.3-FlashX is a dedicated high-speed iteration built on that foundation to target maximum throughput and low-latency environments.

GLM-5.3-FlashX inherits the core architectural design of the GLM-5.3 series, featuring 320B total parameters and 18B active parameters. Built on a hybrid mechanism of Sparse Attention and Linear Attention, the model has undergone low-level operator and communication optimizations on domestic 10000-card computing clusters, supporting long context processing up to 1M tokens.

Core Specifications, Pricing, and Benchmark Throughput Comparison

Compared to the base GLM-5.3-Flash, GLM-5.3-FlashX delivers generation throughput of up to 200 tokens/s, representing an approximate 5x efficiency improvement. Zhipu Open Platform has simultaneously opened the Chat Completion API under the model code glm-5.3-flashx.

In terms of pricing strategy, Zhipu adopted a premium-for-speed model, setting the API rate at 2.5 times that of the standard Flash tier. Engineering teams can evaluate the specifications and pricing in the comparison below:

Model VersionPeak Generation SpeedActive / Total ParametersContext WindowInput Rate (RMB / M Tokens)Output Rate (RMB / M Tokens)
GLM-5.3-Flash40 tokens/s18B / 320B1M0.82.8
GLM-5.3-FlashX200 tokens/s18B / 320B1M2.07.0

Additionally, Zhipu confirmed that context caching storage (Prompt Caching) is currently offered under a limited-time free tier, helping teams reduce cumulative costs in multi-turn dialogues and repository-scale document retrieval.

Engineering Trade-offs for High-Concurrency Agent Workflows

The acceleration in GLM-5.3-FlashX directly targets multi-step AI Agent loops and complex codebase refactoring tasks. In scenarios involving multi-turn tool calling, sandbox code execution, and test-driven self-healing loops, 200 tokens/s throughput compresses long waiting intervals to seconds, substantially reducing end-to-end latency.

For engineering teams, the model selection revolves around cost versus immediacy. For non-realtime batch workflows or offline summarization, the economical GLM-5.3-Flash remains highly cost-effective; however, for IDE-based inline completion, multi-agent orchestration, and real-time interactive systems, GLM-5.3-FlashX delivers a markedly smoother response experience.

Next step

Keep tracking GLM-5.3-FlashX

Continue along the same topic.

Open entity record