GLM-5.3-Flash

Z.ai published GLM-5.3-Flash on August 26, 2026, introducing the first natively multimodal model in the GLM-5 series. The model is a 320-billion-parameter mixture of experts that activates 18 billion parameters per token and was trained from a newly designed base rather than post-trained on GLM-5.2. It combines sparse According to the official Z.ai blog, GLM-5.3-Flash scores 57 on the Artificial Analysis Intelligence Index v4.1.1 at roughly $0.045 per task (discounted), a level previously available only at about ten times the cost. It outperforms GLM-5.2 across six coding and agentic benchmarks — for example 63.4 vs 46.2 on DeepSWE

An analysis published September 1, 2026 reports that Z.ai released GLM-5.3-Flash on August 26, 2026, under an MIT license with weights on Hugging Face on day one. The model is a 320-billion-parameter mixture of experts that activates 18 billion parameters per token, is the first natively multimodal entry in the GLM-5 s The article documents that GLM-5.3-Flash is identical to "Ox Alpha," the anonymous model that had been serving traffic on OpenRouter and drawing attention for its 1M-token context. Pricing is set at $0.15 per million input tokens and $0.50 per million output, compared with $1.40 and $4.40 for GLM-5.3. It also notes tha

The Linas Substack reconstructs the six-day anonymous launch arc of ox-alpha on OpenRouter and OpenCode: an unowned listing with no model card, 1M-token context, and text/image/video inputs appeared on August 20, 2026, and by August 26 it had become the most-used model on the platform, having processed roughly 23 trill The Substack also reports the market response to the identity reveal: weights landed on Hugging Face under MIT on the evening of August 26, and Z.ai/Zhipu's Hong Kong-listed shares closed more than 12 percent higher the next day, roughly 10 times the January IPO price. Beyond the release mechanics the piece contextuali

A Medium explainer dated August 27, 2026 describes GLM-5.3-Flash as Z.ai's newest open-weight model with 320 billion total parameters and about 18 billion active parameters, supporting multimodal inputs and a context window of up to one million tokens. Released on August 26, 2026, under an MIT license, the model is des The article frames GLM-5.3-Flash as combining text, image, and video understanding at substantially lower inference cost than larger GLM-5 siblings, positioning it as a strong candidate for high-volume and multimodal agent applications. It reiterates the headline MoE architecture and open-weight availability, summarizi

An August 27, 2026 comparison piece positions GLM-5.3-Flash against the flagship GLM-5.3 on coding, benchmarks, price, and speed. On the Artificial Analysis Intelligence Index, GLM-5.3 scores 60 versus 57 for GLM-5.3-Flash, but Flash costs roughly one-ninth as much at normal API pricing and activates only 18B parameter According to the comparison, GLM-5.3 generates tokens faster at roughly 85 tokens per second versus about 49 for GLM-5.3-Flash, but Flash has a slightly faster time to first token at 1.46 seconds versus 1.57 seconds. API pricing is set at $0.15 per million input and $0.50 per million output for Flash, against $1.40 and

DataCamp's coverage frames GLM-5.3-Flash as Z.ai's cost-optimized sibling to the full GLM-5.3 flagship, assigning it an Artificial Analysis intelligence index of 57 versus 60 for the larger GLM-5.3. The piece reports a Terminal-Bench 2.1 score of 84.3, positioning the Flash model within striking distance of Claude Opus The post reiterates the 320B-A18B MoE configuration, 1M-token context, native multimodality, and the ox-alpha anonymous preview origin story confirmed by Bloomberg on August 26. Beyond benchmark numbers it frames a trade-off matrix: GLM-5.3-Flash reaches near-frontier coding/agentic performance at a fraction of the cos

BenchLM's aggregator profile records GLM-5.3-Flash's release date as August 26, 2026, classifying it as open-weight, reasoning-capable, with a 1M-token context window. The composite capability score is 66.1 out of 100, ranking the model 34th of 232 tracked models, with 18 source-displayable benchmark rows and no compar The profile flags GLM-5.3-Flash's strongest published category as Multimodal and Grounded (rank 10) and notes that infrastructure cost for self-hosting varies while no comparable first-party hosted token rate is published, distinguishing deployment cost from capability ranking. As of the September 4, 2026 snapshot, the

A technical deep dive published on August 26, 2026 details GLM-5.3-Flash's architecture as a 320B-A18B hybrid-attention MoE that runs natively in FP8 and supports a 1,048,576-token context window. It is the first natively multimodal model in the GLM-5 series and ships under an MIT license on Hugging Face, with API pric The deep dive also contextualizes the release against GLM-4.6, GLM-5.2, and GLM-5.3, observing that GLM-5.3-Flash was built on a freshly trained base — not a post-train of GLM-5.2 — and that GLM-5.3's promised 744B open weights were not included in this drop, with Z.ai's Hugging Face organization still lacking a GLM-5.