GLM-5.3-Flash

GLM-5.3-Flash is Z.AI's efficiency flagship from the GLM-5 series, publicly released in early October 2026 under the MIT License. It is a 320-billion-parameter mixture-of-experts model with just 18B active parameters per token, and the first natively multimodal model in the GLM-5 line. Z.AI trained it from a new base rather than post-training an older checkpoint: the architecture combines hybrid sparse and linear attention with Manifold-Constrained Hyper-Connections (mHC) and an IndexPool mechanism that holds a 1M-token context, all on a 30-trillion-token multimodal pretraining corpus. Inference is supported in SGLang, vLLM, Transformers, KTransformers, TokenSpeed, and Unsloth, and the chat template exposes a three-level reasoning_effort control (low, high, max) for the thinking budget.

On Z.AI's own evaluations, GLM-5.3-Flash beats GLM-5.2 across coding and agentic benchmarks while approaching Claude Opus 4.8, and on Artificial Analysis's Intelligence Index v4.1.1 it scores 57 at a fraction of the per-task cost of frontier models. The standout third-party datapoint is Zapier's AutomationBench, where it scores 48.8 against Claude Opus 4.8's 41.0 and GPT-5.6 Terra's 37.2. Treat all vendor-published benchmark numbers as provider-claimed; the cost-per-task advantage is the better-verified part of the story, since the API lists at $0.15/$0.50 per million input/output tokens.