Z.ai Releases GLM-5.3-Flash, a Natively Multimodal MoE with 1M-Token Context

Architecturally, Z.ai said GLM-5.3-Flash starts from a newly trained base model on a 30-trillion-token multimodal corpus and combines linear and sparse attention for the first time in the GLM series. The 45-layer language model interleaves KDA linear-attention layers with NoPE sparse MLA layers, routes each token through 8 of 288 experts, and ships native FP8 weights plus one MTP draft layer. An IndexPool module compresses groups of indexer key vectors through weighted pooling, which Z.ai says reduces attention compute by about three times and shrinks the KV cache to 4.4 times smaller than GLM-5.3. The model also adopts Manifold-Constrained Hyper-Connections; against GLM-4.5 at similar total parameter count, Z.ai said the model roughly halves both activated parameters and layer count.

On Z.ai-reported benchmarks, GLM-5.3-Flash scores 84.3 on Terminal-Bench 2.1, versus 85.0 for Claude Opus 4.8 and 87.4 for GPT-5.6 Terra, and 63.4 on DeepSWE v1.1, ahead of GLM-5.2's 46.2. It scores 48.8 on AutomationBench, 55.3 on HLE, 62.4 on OfficeQA Pro (ahead of Opus 4.8), and 29.0 on Z.ai Code Bench v1.0, just below Opus 4.8's 29.5. Independently, Artificial Analysis gives it a score of 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, calling it a strong intelligence-per-dollar option but slow and verbose. Vision is the weaker area, with the model trailing Gemini 3.7 Flash on BabyVision and MVbench.

Z.ai said the entire Ox Alpha preview ran on domestically produced Chinese AI chips using a custom SGLang-based engine that disaggregates encoding, prefill and decoding, and reported a 3x end-to-end serving improvement across tens of thousands of accelerators. Standard API pricing is $0.15 per million input tokens, $0.03 per million cached input tokens and $0.50 per million output tokens. The model is included in all GLM Coding Plan tiers — Lite at $18 per month, Pro at $80 and Max at $168 — with three times the usable quota of GLM-5.3. Its multimodal capabilities are surfaced in ZCode through Browser Use and Computer Use; local serving is supported on SGLang, vLLM, TokenSpeed and KTransformers.

For organizations considering self-hosting, the default FP8 checkpoint is roughly 306 GiB of weights before KV cache, and the current vLLM path supports NVIDIA Hopper and newer only. Z.ai recommends at least an 8-GPU node, or a GB200 tray at TP4, putting self-hosting within reach of mid-size and large organizations and AI-native startups renting GPU capacity; everyone else is expected to consume the model through the hosted API. Immediate fits include software and devtools, IT/BPO automation, financial services and insurance document operations, enterprise BI and back-office knowledge work, e-commerce, and teams shipping UI at volume. Representative applications include repository-scale coding agents, terminal and browser/computer-use agents, million-token log and contract analysis, UI regression checking from screenshots, and spreadsheet/deck/dashboard reasoning.