GLM-5.3 Flash

Z.ai announced GLM-5.3-Flash on August 26, 2026 as the first natively multimodal model in the GLM-5 series, built from a newly trained base rather than a post-train of GLM-5.2. It is a 320-billion-parameter mixture-of-experts that activates only 18 billion parameters per token, and it introduces a hybrid architecture c For serving efficiency, the GLM-5.3-Flash design roughly halves both the activated parameter count (18B vs GLM-4.5's 32B) and the number of layers (45 vs 92) compared to the GLM-4.5 series, while using a hybrid attention scheme to minimize long-context attention costs. The post also notes that Z.ai's full flagship GLM-

Fello AI's technical summary confirms Z.ai's GLM 5.3 Flash release on August 26, 2026 as a 320B-total / 18B-active MoE built on a newly trained base — the first natively multimodal model in the GLM-5 series and the first built on a new base since GLM 5.2. It uses a hybrid sparse and linear attention design with a 1M-to On pricing, the article reports API rates of $0.15 per million input tokens and $0.50 per million output for GLM 5.3 Flash, against $1.40 and $4.40 for the full GLM 5.3 flagship. It notes that Z.ai's promised full GLM 5.3 (744B-class) weights had not appeared at the Hugging Face organization, and provides critical anal

AI Profit Boardroom's guide corroborates Z.ai's GLM 5.3 Flash release on August 26, 2026 as a 320B-total / 18B-active MoE with a 1M-token context window, natively multimodal inputs (text, images, video), MIT-licensed open weights, and an anonymous preview phase as Ox Alpha that Z.ai publicly confirmed in its announceme The guide frames free access paths: self-hosting via the MIT-licensed open weights, or using Z.ai's launch discount that halves the already-low API price. Z.ai positions GLM 5.3 Flash at roughly one-tenth of the flagship GLM 5.3 API rate (or one-twentieth during the limited discount), targeting efficient coding and lon

A Medium explainer dated August 27, 2026 describes GLM-5.3-Flash as a 320B-total / 18B-active open-weight MoE from Z.ai with multimodal inputs and a 1M-token context window, released under the MIT license. It confirms the Ox Alpha origin story: before Z.ai's official reveal, developers encountered an anonymous listing The article frames GLM-5.3-Flash as Z.ai's play on the intelligence-per-compute axis, targeting frontier-level coding and agentic capabilities without frontier-level inference costs. It highlights multimodal capabilities spanning text, image, and video understanding as the model's distinguishing feature within the GLM-

DataCamp reports that Z.ai released GLM-5.3-Flash as the cost-optimized sibling to its flagship GLM-5.3, with a 320B-A18B MoE design and 1M-token context window. It notes the model was previewed anonymously under the codename "Ox Alpha" before the public reveal. Z.ai ships GLM-5.3-Flash as natively multimodal with an M DataCamp highlights that GLM-5.3-Flash scores 84.3 on Terminal-Bench 2.1, within striking distance of Claude Opus 4.8 (85.0) and GPT-5.6 Terra (87.4), with an Artificial Analysis Intelligence Index of 57 (versus 60 for GLM-5.3). The blended price is roughly $0.10 per million tokens, about a tenth of GLM-5.3's $0.90, wi

The alphaXiv preprint page for GLM-5.3-Flash introduces the model as the first natively multimodal entry in the GLM-5 series, with 320B total parameters and 18B active, a hybrid sparse+linear attention architecture, Manifold-Constrained Hyper-Connections, and training on a 30T-token multimodal corpus. It reports a 57 s Standard Z.AI API pricing is listed as $0.15 input / $0.50 output / $0.03 cached per million tokens, with weights on Hugging Face under MIT and availability across official platforms (Chat, Code Plan, AutoClaw). The page was published about six days before the current date, making it timely technical background for dev

An independent technical deep dive dated August 26, 2026 confirms GLM-5.3-Flash as a 320B-parameter MoE that activates only 18B per token, runs natively in FP8, supports a 1,048,576-token context window, and is the first natively multimodal model in the GLM-5 series. The piece independently corroborates that the model The deep dive frames GLM-5.3-Flash as a different model trained from a new checkpoint on a different corpus rather than a post-train of GLM-5.2's 744B base, and notes the strategic detail that the flagship GLM-5.3 (744B) had launched on August 14, 2026 with open weights promised "in about two weeks," yet Z.ai's Hugging

The llm-stats aggregator ranks GLM-5.3-Flash 18th on its composite LLM Stats Score and breaks down its standing across categories: Chat 3 of 117 (Top 10%), Tool Calling 7 of 194, Reasoning 17 of 363, Vision 24 of 208, Coding 32 of 267, Math 38 of 327, Legal 70 of 211, and Finance 86 of 228. On cost efficiency the model Individual benchmark scores reported include GDPval-AA 1773.00 out of 3000 (rank 2, methodology GDPval-AA v2 Elo, evaluated by Artificial Analysis), CharXiv-R 0.89 out of 1 (rank 8, with temperature 1.0 / top_p 0.95 and 256K context), and Terminal-Bench 2.1 results sourced from z.ai, alongside other evaluation figures.