Z.ai Releases GLM-5.3-Flash: Approaching Claude Opus 4.8 with 18B Active Parameters

Z.ai released GLM-5.3-Flash, the first native multimodal model in the GLM-5 series. With only 18B of the total 320B parameters active, it demonstrates superior performance over the previous GLM-5.2 in benchmarks and real-world workloads while costing one-tenth as much. In coding and agentic benchmarks, it recorded results approaching Claude Opus 4.8.

Efficient Architecture

GLM-5.3-Flash introduces a hybrid architecture combining sparse and linear attention to significantly reduce long-context serving costs. It also applies Manifold-Constrained Hyper-Connections (mHC) to improve scaling efficiency. Compared to the GLM-4.5 series, the total parameter count is similar (320B vs 355B), but active parameters (18B vs 32B) and layer count (45 vs 92) are nearly halved, optimizing for ultra-low-cost inference.

Performance and Cost Competitiveness

It scored 57 points on the Artificial Analysis Intelligence Index v4.1.1, providing an intelligence level that previously cost about 10 times more for $0.045 per token (with discount applied). It consistently outperformed GLM-5.2 across six coding and agentic benchmarks, showing significant gaps particularly in DeepSWE v1.1 (63.4 points vs 46.2 points) and AutomationBench (48.8 points vs 26.2 points).

Pre-release Testing and Infrastructure

Before release, it was tested under the anonymous name ox-alpha on OpenCode and OpenRouter, with all traffic processed on Chinese AI chips. To support a 1M token context length, it introduced IndexPool technology to reduce indexer latency and memory overhead.