GLM 5.3 FlashX

Z.ai's API lists GLM-5.3-FlashX as a faster hosted variant that is not among the downloadable open-weight models, a material distinction for anyone self-hosting the family. The same guide details the underlying GLM-5.3-Flash as a 320B total, 18B active mixture-of-experts model with 45 layers, 34 linear-attention and 11 sparse-attention layers, and 288 routed experts using eight per token. For reference, GLM-5.3-Flash weights are 328.3 GB in FP8 and run on four H200-class GPUs, or as a community 4-bit build on a 256 GB Mac, with a 1M-token context and up to 128K output. It is MIT licensed and its first natively multimodal design in the GLM-5 series, with reasoning always on at low, high, or max effort.

GLM-5.3-FlashX was released on September 18, 2026 with a published peak inference speed of 200 tokens per second, and it is not a separate model but the same GLM-5.3-Flash system served through a faster inference configuration. The API model ID is glm-5.3-flashx, sitting alongside glm-5.3-flash in Zhipu's lineup. Because the underlying system is unchanged, FlashX carries 320B total and 18B activated parameters, a 1M-token context window, up to 128,000 output tokens, and native image, video, file, and text input. Zhipu's list pricing is $0.37 input and $1.25 output per million tokens, above Flash at $0.15 and $0.50, while Flash is the tier with open weights and included in the GLM Coding Plan.

Z.ai released GLM-5.3-Flash on August 26, 2026, with native vision and a 1M-token context window. The model accepts text, images, video, and files, produces text, and operates in thinking mode only. Z.ai reports that it outperformed GLM-5.2 across six coding and agentic benchmarks, including scores of 63.4 on DeepSWE v1.1 and 48.8 on AutomationBench. GLM-5.3-Flash combines linear attention for local dependencies with sparse attention for global context. Z.ai says this reduces attention compute by 3.01 times and KV-cache size by 4.44 times versus GLM-5.3. The model has 320B total parameters, activates 18B, and targets coding workflows such as screenshot-to-Next.js reproduction, game prototypes, CAD models, and computer use.

Z.ai introduced GLM-5.3-FlashX as a high-speed inference variant of its natively multimodal GLM-5.3-Flash, delivered as a live API tier aimed at agentic coding throughput and complex multimodal workloads where latency matters. The tier keeps the base model's core capabilities while emphasizing faster response for real-time applications. The underlying GLM-5.3-Flash uses a hybrid sparse and linear attention architecture with 320 billion total parameters and 18 billion activated, supporting a large context window plus multimodal input. The base model was released under an MIT license on Hugging Face, and FlashX was previewed anonymously under the code name Ox Alpha before launch.

Benchable.ai lists GLM-5.3-FlashX as the high-speed variant of Z.ai's GLM-5.3-Flash, released September 18, 2026, with a 1M context window and native multimodal input covering image, video, and text, producing text output. The page notes the variant inherits the same hybrid sparse and linear attention architecture as the base Flash model. The model supports tools, top-p, temperature, response format, tool choice, reasoning and max tokens, and is scored at 100% success across eight benchmarks. Reported accuracies include coding at 97.0%, non-hallucination at 98.0%, instruction following at 78.0%, and ethics at 99.0%, with Z.AI listed as the endpoint provider.

Z.AI's developer documentation describes GLM-5.3-Flash/GLM-5.3-FlashX as the first native multimodal model in the GLM-5 series, with 320B total parameters and 18B activated, and is also fully available on the GLM Coding Plan with three times the quota for coding work. It is described as stronger than GLM-5.2 at low cost. The FlashX variant is documented as delivering roughly 200 tokens per second, with 1M-token context, 128K maximum output tokens, and input across video, image, text, and file. The model code glm-5.3-flashx is available through the chat completion API, with text parameters matching GLM-5.3 and recommended settings of temperature 1, top_p 0.95, and reasoning effort max.