Zhipu AI Releases GLM-5.3-Flash: First Multimodal Model in GLM-5 Series, 320B Parameters with Only 18B Active

Zhipu AI Ships GLM-5.3-Flash, a Multimodal MoE Model at 320B Parameters

Zhipu AI (zai-org) has released GLM-5.3-Flash, described as the first natively multimodal model in its GLM-5 series. The model uses a mixture-of-experts (MoE) design with 320 billion total parameters and 18 billion active parameters per forward pass, and weights are available on Hugging Face in BF16, F8_E4M3, and F32 tensor formats.

According to Zhipu AI, GLM-5.3-Flash outperforms its predecessor, GLM-5.2, across benchmarks and real-world workloads "at one-tenth the price," and approaches Claude Opus 4.8 on coding and agentic benchmarks. No specific benchmark score tables or dollar pricing figures were included in the release materials — these are company claims pending independent verification. Pricing is not yet disclosed.

Architecture Changes

GLM-5.3-Flash starts from a newly trained base model rather than a fine-tune of prior GLM checkpoints. Zhipu AI states the architecture introduces two notable changes for the GLM series:

• Hybrid sparse and linear attention, aimed at cutting long-context serving costs while preserving long-context accuracy.
• Manifold-Constrained Hyper-Connections (mHC), a technique the company says improves scaling efficiency during training.

The model was pre-trained on a 30-trillion-token multimodal corpus, larger and more varied than prior GLM training runs, which Zhipu AI credits for the efficiency gains.

Evaluation Notes

The release documentation lists evaluation protocols across several benchmarks, including HLE (with tools), NL2Repo, DeepSWE, Terminal-Bench 2.1, Agent's Last Exam (Toolathlon Verified), AutomationBench, GDPval-AA v2, and a vision test called BabyVision. Evaluations used context windows ranging up to 300,000 tokens (HLE) and up to 1 million tokens (NL2Repo), with some runs using GPT-5.6-luna as a judge model. No numeric scores for these benchmarks were published in the model card, so direct comparisons to GLM-5.2 or Claude Opus 4.8 cannot yet be independently confirmed.

Deployment

GLM-5.3-Flash supports local and self-hosted deployment through several inference frameworks: SGLang, vLLM, TokenSpeed, and KTransformers. It's also available via the Z.ai API Platform and through Hugging Face Inference Providers. The model card lists one fine-tune and two quantized variants already available on the Hugging Face model tree.

The work accompanies a technical report, "GLM-5: from Vibe Coding to Agentic Engineering," posted to arXiv (2602.15763) with a large multi-author list from the GLM-5 team, Zhipu AI, and Tsinghua University collaborators.

What This Means

GLM-5.3-Flash continues the trend of large MoE models activating a small fraction of total parameters to control inference cost — 18B active out of 320B total puts it in similar territory to other efficiency-focused MoE releases from Chinese labs this cycle. The addition of native multimodality to the GLM-5 line signals Zhipu AI is closing feature gaps with frontier multimodal models from OpenAI, Google DeepMind, and Anthropic.

The claims of matching Claude Opus 4.8 on coding and agentic benchmarks at a fraction of the cost are significant if verified, but the model card provides no concrete benchmark tables, pricing, or context window specification — only descriptions of evaluation methodology. Until third-party benchmark runs and official Z.ai API pricing are published, treat the performance and cost claims as unverified vendor statements rather than confirmed results.