GLM-5.3-Flash makes multimodal AI cheaper
Z.ai has released a 320B-parameter open model with 18B active parameters, native vision and a claimed one-million-token context window.
The 60-second version
GLM-5.3-Flash is an MIT-licensed 320B/18B open model with native vision and a claimed one-million-token context.
Key points
- The model has 320B total parameters and 18B activated parameters per token.
- It accepts image and text inputs natively and is designed for visual work such as code, charts and documents.
- Sparse plus linear attention targets lower long-context compute and KV-cache cost.
- Z.ai's benchmark, price and Chinese-chip serving results are self-reported and need independent reproduction.
Verdict. The release is technically significant for open multimodal deployment, but its real cost-performance position will depend on reproducible tests and practical hardware availability.
Z.ai has released GLM-5.3-Flash, a new open model that puts two ideas together: it can reason over images as well as text, and it is designed to keep long-context inference economical. The model has 320 billion total parameters, but only 18 billion are activated for a token, according to the model card and official documentation.
The releaseA large model with a smaller active path
The model is released under the MIT License on Hugging Face. Its mixture-of-experts design contains 320B total parameters and 18B activated parameters. That distinction matters operationally: the total number describes the model's capacity, while the active count is closer to the amount of expert computation used for each token.
Why it mattersVision is part of the model, not an add-on
GLM-5.3-Flash is described as the first natively multimodal model in the GLM-5 series. The API documentation supports image inputs alongside text, which makes the model relevant to coding screens, charts, documents, dashboards and other visual work where translating everything into words adds friction.
The architectureSparse and linear attention target long context
The model combines sparse attention with linear attention. Z.ai says this hybrid design reduces attention computation and KV-cache size compared with GLM-5.3, while preserving long-context capability. The configuration exposes a maximum position length of 1,048,576 tokens and uses an IndexPool mechanism to compress indexer keys.
The evidence boundaryCompetitive numbers still need independent tests
The official blog also says the model was previewed anonymously as Ox Alpha on OpenCode and OpenRouter, with traffic served on Chinese AI chips. Z.ai describes a dedicated SGLang-based serving stack and reports a threefold end-to-end serving improvement on the same hardware. These are useful clues about engineering direction, but the public evidence is still primarily the company's own account.
What to do with itOpen weights change the deployment conversation
The practical choice is not simply whether GLM-5.3-Flash is the best model. It is whether native vision, a very long context and an MIT-licensed checkpoint fit the workload well enough to justify the hardware and serving complexity. SGLang, vLLM, TokenSpeed and KTransformers are listed as deployment paths, giving operators several ways to test that question.
The important release is not just a bigger model. It is an attempt to make frontier-level multimodal capability easier to run, inspect and deploy outside a closed API.