What it does
GLM-5.3-Flash is introduced on its model card as the first natively multimodal model in the GLM-5 series, with 320B total parameters and 18B active. The publisher says it outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth of the price, while approaching Claude Opus 4.8 on coding and agentic benchmarks. These are the publisher's own claims. It starts from a newly trained base model. The card describes a hybrid of sparse and linear attention meant to cut long-context serving cost, Manifold-Constrained Hyper-Connections for scaling efficiency, and a 30T-token multimodal pre-training corpus. The thinking budget is set with a reasoning_effort parameter that accepts low, high or max (default max), and chat use should pass clear_thinking=true. The weights are hosted on Hugging Face in Safetensors format. The card lists local serving with SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth, and names Z.ai, Together AI, Novita, Fireworks, Featherless AI, DeepInfra and Baseten as inference providers.
Capabilities
- Natively multimodal, with image-text-to-text input
- 320B total and 18B active parameters
- Hybrid sparse and linear attention aimed at lower long-context serving cost
- Reasoning effort set to low, high or max
- Local serving with SGLang, vLLM, TokenSpeed, Transformers, KTransformers and Unsloth
- Available through several inference providers
When to use it
- You want a multimodal model whose weights you can download and serve yourself
- You prefer to call a model through an inference provider rather than host it
- You need to trade off reasoning depth and latency with an effort setting
