GLM-5.3-Flash

Z.ai's first natively multimodal model in the GLM-5 series (image, video, and file input; text output) and its first GLM-5 model built on a newly trained base rather than the 744B GLM-5 checkpoint shared by GLM-5.2 and GLM-5.3. It is a 320B-parameter MoE with 18B active (config.json: 45 layers, hidden size 4096, 288 routed experts plus one shared, 8 routed experts per token, 154,880-token vocabulary; ~321.3B by safetensors count) and the first GLM to use a hybrid linear + sparse attention stack — 34 linear-attention layers interleaved 3:1 with 11 DeepSeek Sparse Attention layers — together with Manifold-Constrained Hyper-Connections (mHC) and a new IndexPool that pools four sparse-attention indexer keys into one for 1M-token contexts. Versus the 355B GLM-4.5 it nearly halves active parameters (18B vs 32B) and layers (45 vs 92); versus GLM-5.3 it cuts attention compute 3.0× and KV cache 4.4×. Context is 1M tokens (1,048,576 positions) with 128K max output, and the model was pretrained on Z.ai's latest 30T-token multimodal corpus; Z.ai reports the GLM-5.3-Flash-Base checkpoint at MMLU 88.1 and LiveCodeBench-Base 37.6 versus GLM-5-Base's 88.3 and 34.4.

Z.ai reports it beats GLM-5.2 across its six-benchmark coding/agentic suite while approaching Claude Opus 4.8 at roughly one-tenth GLM-5.2's price: Terminal-Bench 2.1 84.3 (GLM-5.2 81.0), DeepSWE v1.1 63.4 (46.2), AutomationBench v1.0.6 48.8 (26.2), Toolathlon Verified 78.4 (59.9), Agents' Last Exam 26.3 (20.4), HLE with tools 55.3 (54.7), and GDPval-AA v2 1773 (1504; Opus 4.8 1582). On Z.ai Code Bench v1.0 at max effort it scores 29.0 versus Opus 4.8's 29.5. Vision results include OfficeQA Pro 62.4, CharXiv Reasoning with tools 89.4, Chartography with tools 78.0, BabyVision 53.4, MVBench 77.8, and MMVU 80.5. It scores 42 on the Artificial Analysis Intelligence Index v4.3 at $0.15/$0.50 per million input/output tokens, a level of intelligence Z.ai notes was previously available only at about 10× the cost. Thinking cannot be disabled; reasoning_effort accepts low/high/max (default max).

Z.ai's blog confirms that before launch it tested the model anonymously as "ox-alpha" on OpenRouter and OpenCode, where it became the most popular model of the week with all traffic served on Chinese AI chips (Bloomberg reported Z.ai as the maker on 2026-08-26). Serving uses a purpose-built SGLang-based engine with W8A8 weights, hybrid INT8/FP8/BF16 cache quantization, and Encode–Prefill–Decode disaggregation across tens of thousands of domestic accelerators, reaching per-token cost Z.ai says is comparable to NVIDIA GPUs. Weights ship on HuggingFace in FP8 (GLM-5.3-Flash) and BF16 (GLM-5.3-Flash-BF16) under the MIT license; Z.ai says the recipe is "already shaping our next frontier model."