Z.ai Releases GLM-5.3-Flash: A 320B-A18B Natively Multimodal MoE With a 1M-Token Context
Z.ai has launched GLM-5.3-Flash, marking a notable step in the GLM-5 model family by introducing native multimodal capability for the first time. The model uses a Mixture-of-Experts design with 320 billion total parameters, of which only 18 billion are active during any given forward pass - a setup that keeps inference costs lower than a dense model of comparable capacity would require.
The context window sits at 1,048,576 tokens, putting it among the longer-context open models currently available. A key part of what makes that window practical is the attention architecture: GLM-5.3-Flash uses a hybrid scheme combining KDA linear attention with NoPE sparse Multi-head Latent Attention (MLA). Compared to GLM-5.3, this approach cuts attention compute by roughly three times and reduces the KV cache footprint by 4.4 times - meaningful savings when processing very long sequences.
On benchmarks, the model scores 84.3 on Terminal-Bench 2.1 and 63.4 on DeepSWE v1.1, the latter being a coding and software engineering evaluation. These results position it as a capable option for agentic and developer-facing tasks, though as always, benchmark performance does not capture every real-world use case. Weights are available on Hugging Face under the permissive MIT license, which allows commercial use without additional restrictions.
Pricing through the Z.ai API is set at $0.15 per million input tokens and $0.50 per million output tokens, placing it in a competitive range alongside other efficient open-weight models. The combination of a large total parameter count, low active parameter count, a million-token context, and open weights makes GLM-5.3-Flash an interesting option for teams working on long-document processing, multimodal pipelines, or extended agentic workflows who want to self-host or access the model via a low-cost API.
