Károly Zsolnai-Fehér presents GLM 5.3 and GLM 5.3 Flash as open-weight modelsAn open-weight AI model makes trained parameter files available under stated terms so organizations can inspect, host, adapt, or operate the model themselves. capable of generating code, images and 3D scenes. He says their strongest experiments approach leading closed models on some benchmarks, although he does not treat that as equal performance across every task.
The architecture uses 320 billion parameters while activating only about 5 percentAn active AI model parameter is a model weight selected to participate in computation for a particular token or inference step. for each token. It also combines sparse attentionSparse attention reduces transformer computation by allowing each token to attend to a selected subset of other tokens instead of comparing with every token. with cheaper linear attentionLinear attention reformulates attention so its computation grows roughly linearly with sequence length rather than quadratically like standard full attention., plus an indexed memory mechanism that compresses long context before retrieval. Together, those choices reduce the computation neededInference cost is the expense of running a trained AI model to produce outputs for real requests or workloads. for long sessions and large code or document collections.
Károly Zsolnai-Fehér cautions that running the full model locally still requires hardware costing thousands of dollars. Smaller quantized versions are more accessible but can loop or degrade, so the practical opportunity is inexpensive experimentation rather than effortless frontier performance on ordinary consumer machines.
Watch on YouTube
