GLM-5.3-Flash opens with MIT weights, native multimodality, and Flash-tier pricing (2026)

Z.ai released GLM-5.3-Flash on August 26, 2026, calling it the first natively multimodal model in the GLM-5 series. The company published MIT-licensed weights on Hugging Face and opened hosted access under the API model ID glm-5.3-flash. Z.ai describes a newly trained 320B-parameter Mixture-of-Experts model with 18B active parameters, a 1M-token context window, and company-reported coding and agent scores that approach Claude Opus 4.8 while pricing well below GLM-5.3. 1 2 3

What launched

Signal	Confirmed detail	Why it matters
Release	Official research blog dated 2026-08-26; Hugging Face repo `zai-org/GLM-5.3-Flash` last modified 2026-08-26T13:50:10Z with MIT license and safetensors weights. 1 2	Hosted and self-host paths open on the same day.
Architecture	320B total / 18B active parameters; hybrid sparse and linear attention plus Manifold-Constrained Hyper-Connections; max position embeddings of 1,048,576 tokens in the published config. 1 2	A new base, separate from the earlier text-only GLM-5.3 post-training upgrade.
Multimodal API	Image inputs use `type: image_url` content blocks; recommended settings include `temperature: 1`, `top_p: 0.95`, and `reasoning_effort: max`, with thinking required. 3	Screenshot and UI feedback can enter the same coding/agent loop as text.
Price	List API price is $0.15 input / $0.50 output per 1M tokens; a 50% promotion runs through 24:00 on September 9, 2026 (UTC+8) at $0.075 / $0.25. GLM-5.3 remains $1.4 / $4.4. Coding Plan users get 3× the usable quota of GLM-5.3. 1 4	Cost is the clearest product differentiator versus the flagship tier.
Local serving	Official cookbooks list SGLang, vLLM, TokenSpeed, and KTransformers. 2	Self-hosting is documented, though a 320B MoE still needs substantial hardware.

What the company-reported scores show

Z.ai bar chart comparing GLM-5.3-Flash with GLM-5.2, DeepSeek-V4-Vision-Exp, Claude Opus 4.8, GPT-5.6 Terra, and Gemini 3.7 Flash across coding and agent benchmarks — Z.ai's launch chart reports Terminal Bench 2.1 at 84.3, DeepSWE v1.1 at 63.4, AutomationBench v1.0.6 at 48.8, and Agents' Last Exam at 26.3 for GLM-5.3-Flash, with footnotes for harness and sampling settings. 1

On Z.ai's table, GLM-5.3-Flash beats GLM-5.2 by wide margins on DeepSWE (63.4 vs 46.2) and AutomationBench (48.8 vs 26.2) and sits near Claude Opus 4.8 on several coding and agent rows. Vision rows such as OfficeQA Pro (62.4) and Chartography with tools (78.0) are part of the same company table. These figures are Z.ai-reported unless a third party reruns them under the same harness. 1

Before the named release, Z.ai says it ran the model anonymously as ox-alpha on OpenCode and OpenRouter and served that traffic on Chinese AI chips. The blog claims a 3× end-to-end serving gain on that domestic-chip stack versus its own baseline. 1

Why this matters

GLM-5.3-Flash is a different product from the August 14 GLM-5.3 text flagship: new base, native vision, open MIT weights on day one, and API list pricing about one-tenth of GLM-5.3's. For teams already on the GLM Coding Plan or evaluating open multimodal coding agents, the immediate path is a small hosted trial with screenshots and tool loops, plus a hardware check before downloading the full MoE checkpoint. Independent benchmark reruns and real workload latency will decide whether the company table holds outside Z.ai's harness.