Z.ai Releases 320B GLM-5.3-Flash With MIT Weights, 1M Context and Low API Rates
The model gives developers a permissively licensed route to long-context, image and video workloads, while its performance and serving-efficiency claims remain company-reported results.
Listen to this story
The audio brief
Story brief
3 key pointsZ.ai’s GLM-5.3-Flash expands the company’s open-model strategy beyond text, pairing native image/video handling with a 1-million-token context window and permissive MIT weights. Its sparse mixture-of-experts design activates 18B of 320B parameters per token, while API pricing starts at $0.15 per million input tokens. The release could appeal to developers needing long-context multimodal systems at low serving cost,...
- 01
The model is available in FP8 and BF16 on Hugging Face and ModelScope under the MIT license.
- 02
API pricing is $0.15 per million input tokens, $0.50 output, and $0.03 for cached input.
- 03
GLM-5.3-Flash scored 63.4 on DeepSWE v1.1, versus 46.2 for GLM-5.2.
Z.ai has released GLM-5.3-Flash, a 320-billion-parameter mixture-of-experts model that activates 18 billion parameters for each token. It combines MIT-licensed open weights, native image and video input, a 1-million-token context window and API input priced at $0.15 per million tokens.
The August 26 release is the first natively multimodal model in Z.ai’s GLM-5 series. Its mixture-of-experts design has 320 billion total parameters but selects 18 billion per token. Z.ai says the model supports image and video inputs natively, rather than limiting the release to text.
GLM-5.3-Flash uses a hybrid sparse- and linear-attention architecture. Z.ai describes that design as cutting compute requirements significantly, a consequential claim for a model built to accept up to 1 million tokens of context in one request.
Z.ai lists input usage at $0.15 per million tokens.
Z.ai lists output usage at $0.50 per million tokens.
Cached inputs are listed at $0.03 per million tokens.
The company offers the model through its API and has released the weights under the MIT license in FP8 and BF16 formats on Hugging Face and ModelScope. The permissive license allows use, modification and commercial distribution of the weights, according to the launch coverage.
On the DeepSWE v1.1 coding benchmark, GLM-5.3-Flash scored 63.4, compared with 46.2 for GLM-5.2. That establishes a reported improvement over the prior model; it does not by itself establish a ranking against other frontier systems.
Z.ai says GLM-5.3-Flash runs across clusters of domestically manufactured Chinese AI accelerators. It also reports that custom serving stacks and hardware-tailored quantization produced a threefold end-to-end performance improvement over a more generic deployment approach. How that result transfers to other hardware and workloads remains unresolved.
Editorial analysis
Our Read
Our read: the strategic value is the package, not a single score. GLM-5.3-Flash combines MIT-licensed weights, a 1-million-token window and listed API rates that give teams a choice between hosted access and adapting the released model. That places it alongside a growing set of low-cost, long-context multimodal releases, including Qwen3.8-Flash. The next meaningful test is whether users can reproduce the reported coding and serving gains on workloads that matter to them, especially across deployment stacks outside Z.ai’s hardware-tailored configuration.
Sources
- cryptobriefing.comZhipu launches GLM-5.3-Flash, its first natively multimodal model built for Chinese chips
