320B total / 18B active (MoE; official checkpoint contains approximately 321.3B parameters)
1,048,576 tokens
MIT
Official open-weight release of GLM-5.3-Flash, comprising native FP8 and BF16 checkpoints and confirming the identity of the anonymously tested ox-alpha stealth model. GLM-5.3-Flash is the first natively multimodal model in the GLM-5 series and is distinct from the full GLM-5.3 flagship: whereas GLM-5.3 continues post-training from the GLM-5.2 base, Flash begins from a newly trained base model with a redesigned architecture and a 30-trillion-token multimodal pre-training corpus. The model contains approximately 320 billion total parameters with 18 billion activated per token, versus 744B-A40B for GLM-5.2. Its 45-layer, 4,096-dimensional language backbone generally follows a repeating pattern of three linear-attention layers followed by one sparse-attention layer, with 34 linear and 11 sparse layers in total. The first three feed-forward layers are dense; subsequent layers use 288 routed experts with eight selected per token plus one shared expert. Architectural additions include four-branch Manifold-Constrained Hyper-Connections, one Multi-Token Prediction layer, and IndexPool, which compresses groups of four sparse-attention indexer keys through weighted pooling. Z.AI reports that the hybrid linear-and-sparse attention design reduces attention computation by 3.0 times and average KV-cache size by 4.4 times relative to the full GLM-5.3 at one-million-token context. Native visual processing is provided by a 24-layer, 1,024-dimensional vision encoder with temporal patching, supporting visual coding, document understanding, browser use, computer use, output inspection, and self-verification. The embedded chat template supports low, high, and max reasoning-effort settings, defaults to max, and clears historical reasoning traces unless clear_thinking is set to false. Published evaluations report 84.3 on Terminal-Bench 2.1, 63.4 on DeepSWE v1.1, 56.3 on NL2Repo, 78.4 on Toolathlon Verified, 48.8 on AutomationBench, 26.3 on Agents' Last Exam, and 55.3 on Humanity's Last Exam with tools. Multimodal evaluations include 62.4 on OfficeQA Pro, 89.4 on CharXiv Reasoning with tools, 78.0 on Chartography with tools, 53.4 on BabyVision, 77.8 on MVBench, and 80.5 on MMVU. Z.AI also reports an Artificial Analysis Intelligence Index v4.1.1 score of 57 at a discounted cost of $0.045 per task, describing the model as outperforming GLM-5.2 across benchmarks and real-world workloads at one-tenth the price while approaching Claude Opus 4.8 on coding and agentic benchmarks. The primary FP8 repository contains approximately 314.4 billion FP8 and 6.9 billion BF16 parameters across 62 shards, while the full-BF16 repository contains approximately 321.3 billion BF16 parameters across 120 shards. Local deployment guidance is available for SGLang, vLLM, TokenSpeed, and KTransformers.