A 320-billion-parameter model runs entirely on Chinese AI chips

GLM-5.3-Flash combines multimodal capabilities, a one-million-token context, and MIT-licensed open weights, with 18 billion active parameters.

Running a 320-billion-parameter model for a fraction of the usual price is already a compelling commercial argument. Deploying it at scale without Nvidia hardware adds a more political dimension. With GLM-5.3-Flash, Z.ai is introducing a new multimodal model, an architecture designed to reduce long-context costs, and a demonstration of China’s growing infrastructure independence.

GLM-5.3-Flash is the first natively multimodal model in the GLM-5 family. It accepts text, images, video, and files, then produces text. Its intended uses extend beyond conversation to interface development, application control, document analysis, financial research, presentation creation, and the processing of visual content.

The model has 320 billion parameters but activates only 18 billion for each token. This mixture-of-experts structure selects part of the network according to the request, increasing total capacity without using the entire model at every step.

The difference between total capacity and active computation does not make the model lightweight. The FP8 version published on Hugging Face occupies approximately 328 GB, distributed across 62 main files. The BF16 version reaches 643 GB. “Flash” therefore refers primarily to production efficiency and usage cost, not to a model intended for an ordinary personal computer.

Independent deployment is supported through SGLang, vLLM, TokenSpeed, Transformers, KTransformers, and Unsloth. It still requires infrastructure capable of holding the weights, intermediate states, and context-related cache. Using the full one-million-token window further increases memory requirements, even though all parameters are not activated simultaneously.

That context window can accommodate a long conversation history, a document collection, a substantial portion of a software repository, or numerous visual elements within one request. It does not guarantee that every piece of information will be retrieved with equal accuracy, or that processing will remain as fast as it is with a short input. Technically supporting one million tokens and reasoning reliably across that entire length remain two different measures.

To reduce the cost of these volumes, the model combines linear and sparse attention. The former maintains a compressed representation of previous information instead of systematically comparing each new token with the full context. The latter preserves access to distant information while selecting only the areas considered relevant.

A component called IndexPool further reduces the cost of this global search. It combines four index key vectors into a single weighted representation. According to Z.ai, this compression lowers the latency and memory required to identify useful passages when the context approaches one million tokens.

The model also adopts Manifold-Constrained Hyper-Connections, or mHC. This system expands the routes through which information moves between layers while applying constraints intended to preserve network stability. The goal is to improve scaling without increasing active computation at the same rate.

Compared with the GLM-4.5 series, the number of layers falls from 92 to 45 and the number of active parameters from 32 billion to 18 billion, while total size remains relatively close: 355 billion for GLM-4.5 versus 320 billion here. Much of the Flash designation comes from this shorter, more selective architecture.

Z.ai’s documentation claims a threefold reduction in attention computation and a 4.4-fold reduction in average KV-cache size compared with GLM-5.3. These values are calculated per head and per layer in the configurations being compared. They do not directly translate into an identical acceleration for the complete model, as multimodal encoding, expert routing, memory transfers, and serving infrastructure also affect performance.

Z.ai also acknowledges that the cache remains slightly larger than those of Kimi-K3 and DeepSeek-V4-Flash in its comparison. The model shows the lowest attention-computation requirement among the selected systems, but it does not lead on every component.

Training begins from a new base model and uses a multimodal corpus containing a reported 30 trillion tokens. The company has not disclosed the dataset’s detailed composition, the proportions assigned to each language and content type, or the materials required to reproduce the complete training process. Publishing the weights therefore does not provide access to the entire production pipeline.

The reported results place GLM-5.3-Flash ahead of GLM-5.2 across several coding and agent evaluations. On DeepSWE 1.1, it scores 63.4, compared with 46.2 for its predecessor. It also exceeds the 58 points attributed to Claude Opus 4.8, while remaining behind Gemini 3.7 Flash at 65.3 and GPT-5.6 Terra at 69.6.

On Terminal-Bench 2.1, its score of 84.3 approaches Claude Opus 4.8’s 85, but remains below Gemini 3.7 Flash at 85.8 and GPT-5.6 Terra at 87.4. On AutomationBench 1.0.6, it reaches 48.8, compared with 26.2 for GLM-5.2, 41 for Claude Opus 4.8, and 52.3 for Gemini 3.7 Flash.

These figures need to be considered alongside their execution conditions. Terminal-Bench was run through Claude Code with six hours allowed for each task. DeepSWE uses mini-SWE-agent, a 400,000-token context, and the same time limit. Several evaluations permit tools, very long outputs, or specific context-management methods. The results therefore depend on the model, agent harness, permissions, and selected settings.

Z.ai also presents its internal Code Bench, where GLM-5.3-Flash scores 29 at maximum effort, compared with 29.5 for Claude Opus 4.8. This internal test provides another point of reference but does not replace independent reproduction.

Its visual capabilities include reading documents, charts, interfaces, and videos. The model scores 62.4 on OfficeQA Pro and 78 on Chartography with tools enabled. It can also observe the output of an interface or 3D scene, compare the result with the initial request, and