Z.ai launches GLM-5.3-Flash: first multimodal MoE in series with a million-token context

Z.ai released GLM-5.3-Flash, the first model in the GLM-5 series built from the start as multimodal and also the cheapest the lab has put to code. It uses a Mixture-of-Experts architecture with 320 billion parameters total and 18 billion active per token, a context window of 1,048,576 tokens, and supports image and video input. The weights are available on Hugging Face under an MIT license, and the model spent its first week running anonymously as “Ox Alpha” on OpenCode and OpenRouter, deployed entirely on locally manufactured Chinese AI chips.

**Architecture: hybrid attention and index compression**

Efficiency comes from three main moves. For the first time in the series, Z.ai combines linear attention (KDA) with sparse attention (NoPE sparse MLA) across 45 layers, routing each token through 8 of 288 experts, using native FP8 weights and a single MTP layer for draft. IndexPool compresses key-vector groups of the index via weighted pooling, which the company says cuts attention computation by a factor of 3 and reduces KV cache by 4.4 × compared with GLM-5.3. The mHC (Manifold-Constrained Hyper-Connections) mechanism, against GLM-4.5 with a similar parameter count, halves both active parameters and layer count.

**Deployment: who can run it locally and who will stay on the API**

The default checkpoint in FP8 weighs about 306 GiB before KV cache, and the current vLLM path supports only NVIDIA Hopper and newer. This makes self-deployment feasible for medium- and large-scale organizations with at least an 8-GPU node (or a GB200 board in TP4), as well as AI-native startups that rent GPU capacity. Everyone else will consume the model via API, where economics—not hardware—drive the story. Standard pricing is $0.15 per million input tokens, $0.03 per cached input token, and $0.50 per output token.

**Benchmarks: strong in code, weak in vision**

According to Z.ai, the model beats GLM-5.2 on all benchmarks and real-world workloads at one-tenth the price, and sits just half a point behind Claude Opus 4.8 on the internal coding benchmark. Artificial Analysis ranks it 57 on the intelligence metric v4.1.1, with 48.7 output tokens per second and a TTFT of 1.52 seconds on Z.ai’s API—a strong intelligence-to-dollar ratio, but slower and more verbose. The weak link is vision: the model lags behind Gemini 3.7 Flash on BabyVision and MVbench.

**Serving engine: the engine behind the demo**

The less-publicized part is the serving engine: every run of Ox Alpha was performed on local Chinese chips using a customized SGLang-based engine that separates encoding, profiling, and decoding. Z.ai reports a three-fold end-to-end serving improvement over tens of thousands of accelerators. Immediate adaptation appears in code agents at repo scale, terminal agents and browsers, log and contract analysis at a million-token level, UI regression testing from screenshots, and thinking about spreadsheets/presentations/dashboards that previously required OCR-to-text pipelines.