GLM 5.3 Flash

An anonymous listing called "Ox Alpha" appeared on OpenRouter and OpenCode on August 20, 2026 with a 1M-token context window and text, image, and video input but no model card or owner listed. Within six days, Ox Alpha became the most-used model on the platform, processing approximately 23 trillion tokens, about 2.3 ti The confirmation identifies GLM-5.3-Flash as a 320-billion-parameter model with 18 billion active per token, the first natively multimodal model in the GLM-5 series, and the cheapest model at its score on the Artificial Analysis Intelligence Index. The model ran entirely on Chinese AI chips during its Ox Alpha preview

Z.ai shipped GLM 5.3 on August 14, 2026 as an API-only flagship, then followed with GLM 5.3 Flash twelve days later on August 26, carrying open weights under MIT, native image and video input, and launch pricing roughly a tenth of the flagship's. Z.ai's own launch tables showed Flash matching or beating the larger mode The flagship retains a real edge from a deeper post-training run, with Z.ai publishing a public disclosure ledger crediting 53 CVEs at launch as externally checkable evidence. GLM 5.3 Flash launched with a week of real-world usage behind it from its stealth run as Ox Alpha on OpenRouter, processing roughly 23 trillion

Z.ai's GLM 5.3 Flash is dramatically cheaper than the GLM 5.3 flagship, natively multimodal, and activates only 18B parameters versus 40B for GLM 5.3, while delivering surprisingly close performance. In independent Artificial Analysis testing, GLM 5.3 scores 60 on the Intelligence Index compared with 57 for GLM 5.3 Fla Despite the "Flash" label implying speed, Artificial Analysis measures GLM 5.3 at roughly 85 tokens per second versus about 49 tokens per second for Flash, with GLM 5.3 Flash actually faster only on time-to-first-token (1.46s vs 1.57s). Flash is the only variant with native image input, open weights currently available

A community thread on the NVIDIA Developer Forums reports the release of GLM-5.3-Flash weights on August 26, 2026, linking to a Z.ai organization X post as the upstream source. The model, formerly known as Ox Alpha, is described as a 320B-total / 18B-active parameter MoE with a 1M context window, image+text input, and Follow-on replies on the same thread surface active benchmarking work, including NVFP4 quantization recipes on DGX Spark hardware (e.g., 3x Spark TP=3 at 512K context reaching roughly 35 tok/s), giving concrete run-time evidence alongside the weights release. The signal is community-validated rather than first-party, a

A community PR to the official zai-org/GLM-5.3-Flash Hugging Face repository added evaluation results extracted from the model card's benchmark chart, with a staff member merging the change. The PR introduced Terminal-Bench 2.1 at 84.3, DeepSWE at 63.4, and Humanity's Last Exam at 55.3 as benchmark values from the publ Additional benchmarks from the model card could not be added because their eval.yaml definitions were not registered on the Hugging Face Hub, including Agent's Last Exam at 26.3, AutomationBench v1.0.6 at 48.8, and GDPval-AA v2 at 1773. The extracted evaluation file was committed under .eval results/GLM-5.3-Flash.yaml.

The LLM Stats profile (Aug 26, 2026) ranks GLM-5.3-Flash at #17 overall on its composite LLM Stats Score with a blended price of $0.17 per million tokens and capability tier rankings including Tool Calling 7 of 193 (top 10%), Reasoning 16 of 362, Chat 21 of 117, Vision 24 of 208, Coding 32 of 266, and Math 38 of 327. P The Quality Tracker shows +6.71σ baseline movement and conversation-depth performance (turn 1: 14.3; turns 2-10: 14.6). The page positions GLM-5.3-Flash against Gemma 4 E4B ($0.024), GPT OSS 120B ($0.043), DeepSeek-V4-Flash-0731 ($0.066), Muse Spark 1.3 ($0.10), and GPT-6 Astra ($11.9) on the cost-efficiency curve, tho

This Local AI Zone technical deep dive (Aug 26, 2026) documents GLM-5.3-Flash as a 320B-parameter mixture-of-experts model activating 18B per token, natively running in FP8, with a 1,048,576-token context window, native multimodality (text, image, video input; text output), MIT-licensed weights on Hugging Face, $0.15/$ The deep dive emphasizes that Flash is a different model trained from a newly designed base and training recipe, not a post-train of GLM-5.2's 744B base. It reports that GLM-5.3's promised 744B open-weight release (announced two weeks prior) has not materialized, with no GLM-5.3 repository currently on Z.ai's Hugging F

Z.ai released GLM 5.3 Flash on August 26, 2026, as a 320-billion-parameter mixture-of-experts model that activates 18 billion parameters per token, with MIT-licensed weights published to Hugging Face on day one. The model is the first natively multimodal entry in the GLM-5 series, accepting text, image, and video input The release shifts attention away from the still-unreleased open weights of the 744B GLM 5.3 flagship, and Z.ai's Hugging Face organization had no GLM 5.3 repository at the time of Flash's launch. Within six days of its quiet OpenRouter debut, Ox Alpha had become the platform's most-used model, processing roughly 23 tr