GLM 5.3 Flash

Z.ai officially introduced GLM-5.3-Flash on August 26, 2026, calling it the first natively multimodal model in the GLM-5 series. The model uses a 320B total / 18B active mixture-of-experts design and is built from a newly trained base rather than a post-trained trim of the GLM-5.2 checkpoint. It combines sparse and lin On the Artificial Analysis Intelligence Index v4.1.1, GLM-5.3-Flash scores 57 at roughly $0.045 per task (discounted), pushing the Pareto frontier of intelligence-per-dollar. Across coding and agentic benchmarks it consistently beats GLM-5.2 (DeepSWE v1.1 63.4 vs 46.2; AutomationBench 48.8 vs 26.2) and approaches Claud

Yotta Labs' comparison frames GLM 5.3 (August 14, 2026, API-only flagship) against GLM 5.3 Flash (August 26, 2026, MIT-licensed open weights), highlighting that Flash spent a week running in stealth as "Ox Alpha" before Z.ai claimed it. Flash uses a 320B total / 18B active MoE base with native text, image, and video in The piece warns that the two launch tables use different benchmark versions: the flagship reports Terminal-Bench 3.0 (28.3, a six-fold jump over GLM 5.2's 4.6), while Flash reports Terminal-Bench 2.1 (84.3 vs Claude Opus 4.8's 85.0). The one shared test, DeepSWE v1.1, shows Flash at 63.4, GLM 5.2 at 46.2, and the GLM 5

DataCamp's overview confirms GLM-5.3-Flash as Z.ai's cost-optimized sibling to the GLM-5.3 flagship, released after running anonymously as "Ox Alpha." It uses a 320B-A18B mixture-of-experts design with a 1M token context window and is natively multimodal. The model scores 84.3 on Terminal-Bench 2.1, within striking dis The piece highlights the price-performance trade-off: blended pricing sits at roughly $0.10 per million tokens, about a tenth of GLM-5.3's $0.90, but output throughput is around 49 tokens/sec and the FP8 checkpoint weighs 306 GiB, ruling out lightweight local hosting. Qwen3.8-Flash-Next launched on the same day, intens

Local AI Zone's technical deep dive describes GLM-5.3-Flash as a 320B-parameter MoE activating only 18B per token, running natively in FP8 with a 1,048,576-token context window and shipped under MIT license on Hugging Face. It is the first natively multimodal model in the GLM-5 series and was trained from a new base, n The piece reports that GLM-5.3-Flash beats GLM-5.2 across six coding and agentic benchmarks at roughly one-tenth the price and approaches Claude Opus 4.8 on long-horizon agent tasks. It notes that before launch, "Ox Alpha" ran anonymously on OpenRouter for about a week, topping usage charts with near-unlimited free acc

Z.AI's developer documentation for GLM-5.3-Flash and GLM-5.3-FlashX details a hybrid architecture that reduces attention computation by 3.01× and KV cache by 4.44× versus GLM-5.3, with a 1M context window and 128K maximum output tokens. The model accepts video, image, text and file input and returns text output, and recommended settings are temperature 1, top_p 0.95 and reasoning effort max. Visual capabilities sit inside the coding loop, letting the model observe interfaces and rendered results to test and refine frontend, game and Blender 3D work through browser and computer-use agents, while also handling Office, financial research and document workflows that end in PPTX, PDF, DOCX and XLSX deliverables. GLM-5.3-FlashX is live with inference speeds around 200 tokens per second, and GLM-5.3-Flash is now fully available on the GLM Coding Plan with three times the quota, accessible under the model codes glm-5.3-flash, glm-5.3-flashx and glm-5.3-flashx-legal.

Z.ai has released GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series, with 320B total parameters and 18B active ones. The company says it outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price, while approaching Claude Opus 4.8 on coding and agentic benchmarks, according to the model card. The model starts from a newly trained base with a redesigned architecture and training recipe, combining sparse and linear attention for the first time in the GLM series to sharply cut long-context serving cost, and adds Manifold-Constrained Hyper-Connections for scaling efficiency on a 30T-token multimodal pre-training corpus. It can be deployed on SGLang, vLLM, Transformers, TokenSpeed, KTransformers and Unsloth, with reasoning effort controllable at low, high, or max.