Z.ai Releases GLM-5.3-Flash: A 320B-A18B Natively Multimodal MoE With a 1M-Token Context

Z.ai has put out GLM-5.3-Flash, the first natively multimodal model in its GLM-5 line and, by its own account, the cheapest capable coding model it has shipped. The model is a mixture of experts with 320B total parameters, 18B active per token, image and video input, and a 1,048,576-token context window. The weights are on Hugging Face under an MIT license, and the hosted API is already live.

The company is pushing two very different deployment stories at once. If you want to run it yourself, the barriers are real: the default FP8 checkpoint is about 306 GiB before KV cache, and the current vLLM path is for NVIDIA Hopper and newer. That makes it plausible for mid-size and large orgs with an eight-GPU node or a GB200 tray at TP4, plus startups that already rent plenty of GPU time. Everyone else will use it as an API, which is where Z.ai says the economics get interesting.

That economics claim is tied to performance. Z.ai says GLM-5.3-Flash improves on GLM-5.2 across benchmarks and real workloads at about one-tenth the price, and comes within half a point of Claude Opus 4.8 on the company’s internal coding benchmark. On Z.ai’s reported numbers, it scores 84.3 on Terminal-Bench 2.1 and 63.4 on DeepSWE v1.1, both comfortably ahead of GLM-5.2. Artificial Analysis also puts it at 57 on the Intelligence Index, with 48.7 output tokens per second and 1.52 seconds time to first token on Z.ai’s API.

The architecture is trying to buy that efficiency instead of brute-forcing it. GLM-5.3-Flash was trained from scratch on a 30T-token multimodal corpus, then built with a hybrid attention setup that mixes KDA linear-attention layers and NoPE sparse MLA layers. Z.ai says that cuts attention compute by about 3x and shrinks the KV cache by 4.4x versus GLM-5.3. The model also uses Manifold-Constrained Hyper-Connections, and compared with GLM-4.5 at similar total size, it roughly halves both activated parameters and layer count.

The serving setup may matter as much as the model itself. Z.ai says the first week of the model’s anonymous “Ox Alpha” preview ran on domestically produced Chinese AI chips, using a custom SGLang-based engine. The model is also being sold through GLM Coding Plan tiers at Lite, Pro and Max, with 3x the usable quota of GLM-5.3, and its multimodal features are exposed in ZCode through Browser Use and Computer Use. For teams doing repo-scale coding, long document analysis, UI checks or spreadsheet reasoning, this is a very practical release.