Z.ai has released GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series, featuring a mixture-of-experts architecture with 320 billion total parameters and 18 billion active per token. The model supports image and video inputs within a 1,048,576-token context window and is available under an MIT license on Hugging Face.

  • It employs hybrid attention combining KDA linear-attention layers with NoPE sparse MLA layers, routing tokens through 8 of 288 experts.
  • An IndexPool mechanism reduces attention compute by approximately 3x and shrinks the KV cache by 4.4x compared to GLM-5.3.
  • Benchmarks show scores of 84.3 on Terminal-Bench 2.1 and 63.4 on DeepSWE v1.1, placing it near Claude Opus 4.8.
  • API pricing is set at $0.15/M for input tokens and $0.50/M for output tokens, with self-hosting requiring ~306 GiB of FP8 weights.

The model aims to provide a cost-effective coding solution, reportedly beating GLM-5.2 across benchmarks at roughly one-tenth the price while maintaining high performance on internal coding tasks.