Z.ai Releases GLM-5.3-FlashX, a 200 Tokens/Second Variant of Its GLM-5.3-Flash Model

Z.ai released GLM-5.3-FlashX on September 18, 2026, a high-speed variant of its GLM-5.3-Flash model designed for fast inference in coding, visual understanding, and long-horizon agent workloads.

Key Specifications

GLM-5.3-FlashX is a native multimodal model built on the same hybrid sparse and linear attention architecture that underlies GLM-5.3-Flash. According to Z.ai, the model uses 320B total parameters with 18B active parameters at inference time — a mixture-of-experts-style design intended to cut compute overhead while preserving accuracy over long context windows.

The model supports a 1M-token context window and, according to Z.ai, delivers inference speeds of up to 200 tokens per second. Pricing for GLM-5.3-FlashX has not yet been disclosed; the base GLM-5.3-Flash model it derives from is listed at $0.075 per 1M input tokens and $0.25 per 1M output tokens on OpenRouter, though FlashX's own rate has not been separately confirmed.

Positioning Within the GLM Lineup

GLM-5.3-FlashX sits within Z.ai's broader GLM-5.3 family, which also includes GLM-5.3-Flash (offered in both 1.3M and 1.0M context variants at $0.075/$0.25 per 1M tokens) and GLM-5.3, a reasoning model priced at $0.70-$0.90 per 1M input tokens and $2.20-$2.83 per 1M output tokens depending on context configuration. Unlike GLM-5.3, which keeps reasoning always on, GLM-5.3-FlashX is positioned as a speed-optimized, non-reasoning variant aimed at latency-sensitive coding and agent applications.

Z.ai has been iterating rapidly through its GLM series, moving from GLM-4.5V and GLM-4.6 through GLM-5, GLM-5.1, GLM-5.2, and now GLM-5.3 variants within roughly a year. The FlashX designation follows a pattern seen elsewhere in the industry, where companies ship a lower-latency, cost-reduced sibling alongside a flagship reasoning model to serve high-throughput production use cases.

What This Means

GLM-5.3-FlashX targets a specific gap: applications that need near-instant responses at large context sizes without paying the latency or cost premium of a full reasoning model. The 200 tokens/second claim, if it holds up in independent testing, would make it competitive for real-time coding assistants and agent loops that call models repeatedly within a single task.

The missing piece is pricing. Without a confirmed rate for FlashX specifically, it's difficult to assess whether the speed gains come with a cost trade-off relative to the standard GLM-5.3-Flash tier. The 320B/18B active parameter split suggests Z.ai is using aggressive sparsity to keep serving costs down even at high throughput, which is consistent with its stated goal of balancing performance against token efficiency across the GLM-5.3 generation. Independent benchmarks on coding and agentic tasks will be needed to verify Z.ai's throughput and capability claims.