模型與推論系統
GLM-5.3-Flash Uses Hybrid Sparse/Linear Attention to Cut Million-Token Inference Costs
Zhipu AI has unveiled GLM-5.3-Flash, a native multimodal model with 320B total parameters and 18B active parameters per token, releasing its weights under the MIT License. The new architecture is claimed to require three times less attention computation and 4.4 times less KV cache than GLM-5.3, but “Flash” does not mean it can be easily deployed on a typical workstation.
Zhipu AI has released additional details on GLM-5.3-Flash’s architecture, deployment options, and benchmark results. The model has 320B parameters in total, but its mixture-of-experts path activates only 18B parameters per token. Its layer count has also been reduced from 92 in the similarly sized GLM-4.5 series to 45. Rather than being directly pruned from GLM-5.3, it is a native multimodal foundation model pretrained from scratch on 30T tokens of image-text data, with support for text, images, video, and documents as input.
Long-context processing is the most concrete engineering change in this release. The model interleaves linear attention with sparse attention: the former uses state modeling to preserve local dependencies, while the latter relies on a lightweight indexer to select relevant positions from the global context. When processing up to one million tokens, IndexPool further compresses four cached key vectors into one through weighted aggregation. Zhipu AI says that, compared with GLM-5.3, the model cuts attention computation to roughly one-third and KV cache usage to about one-fourth point four. Its serving stack also uses ReplaySSM, W8A8, a hybrid INT8/FP8/BF16 cache, and disaggregated deployment for encoding, prefill, and decode.
The weights are available on Hugging Face under the MIT License. The model can currently be launched with Transformers, vLLM, SGLang, KTransformers, and TokenSpeed, and offers three levels of control through `reasoning_effort`. The model card warns that the highest reasoning effort is used unless explicitly specified. For chat workloads, `clear_thinking=true` is recommended; otherwise, latency and output behavior in multi-turn deployments may differ from expectations.
Zhipu AI reports scores of 84.3 on Terminal Bench 2.1, 63.4 on DeepSWE v1.1, and 78.4 on Toolathlon Verified, though these remain vendor-reported evaluations. Artificial Analysis shows that generation speeds across API providers can differ by more than tenfold, illustrating that model architecture is only one factor in real-world throughput. Engineering teams should next verify visual and tool-use capabilities after quantization, actual KV memory consumption, and cost-versus-accuracy at equivalent reasoning effort. The model’s 320B total weights also mean that full self-hosting remains a server-class workload.