Z.ai open-sourced a 320-billion-parameter model under the MIT license, claiming a 10x cost cut
GLM-5.3-Flash is natively multimodal with a roughly 1-million-token context window, confirmed directly from its published config -- but Z.ai's own cost-efficiency and benchmark-leadership claims are vendor comparisons, not independently verified numbers.
Z.ai released GLM-5.3-Flash on August 26, 2026: a mixture-of-experts model with 320 billion total parameters and 18 billion active per token, published under the MIT license on Hugging Face. Its own model card describes it as "the first natively multimodal model in the GLM-5 series," and its published configuration file sets a maximum position embedding of 1,048,576 tokens -- a roughly 1-million-token context window, confirmed directly from the model's own config rather than taken from marketing copy.
What's confirmed, and what's Z.ai's own claim
The parameter counts, multimodality, context length, and MIT license all come directly from Z.ai's published artifacts. The cost and performance claims don't: Z.ai's own model card says the model "outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth the price," and SiliconANGLE reports Z.ai claiming the top score among compared models on the GDPval-AA v2 benchmark. Both are Z.ai's own comparisons against its own predecessor and its own choice of benchmark competitors -- worth noting plainly as vendor-claimed rather than independently verified, the same distinction this site draws whenever a lab publishes its own efficiency or benchmark numbers.
Why the license matters as much as the specs
An MIT-licensed, million-token-context model with a claimed order-of-magnitude cost advantage is precisely the kind of release that complicates a company's own AI spend tracking: self-hosting an open-weight model shifts cost from a per-token API line item to compute and ops overhead that doesn't show up the same way on a bill. Whether that shift is actually cheaper for a given workload depends entirely on numbers a spend/value framework has to measure directly -- not on a vendor's own comparison chart, however credible the underlying model turns out to be.