Zhipu AI has released GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series and the first open-weight release of the glm5_next architecture. The 320B-parameter model is trained on a 30T-token multimodal corpus and features a hybrid sparse and linear attention mechanism to reduce long-context serving costs.

  • Architecture: 45 layers combining KDA linear attention with DeepSeek-style sparse attention, utilizing Manifold-Constrained Hyper-Connections (mHC) for scaling efficiency.
  • Multimodal capabilities: A 24-layer ViT encoder allows image and video tokens in the vocabulary, supporting a context length of up to 1,048,576 tokens.
  • Weights: Available in FP8 (e4m3) and BF16 formats on Hugging Face and ModelScope under an MIT license.
  • Inference support: Official recipes provided for vLLM and SGLang, including speculative decoding configurations.

The release aims to offer a cost-effective alternative that approaches the performance of Claude Opus 4.8 on coding and agentic benchmarks while significantly reducing inference costs.