GLM-5.3 Flash (CanopyWave)

A dated independent review identifies the anonymous model previously seen on OpenRouter and OpenCode as Ox Alpha and reports that Z.ai officially released GLM-5.3-Flash on August 26, 2026. It describes the model as a 320B total, 18B-active MoE with 1M-token context, native image and video input, MIT open weights, and support for vLLM, SGLang, and TokenSpeed. Z.ai's reported launch numbers include 84.3 on Terminal-Bench 2.1, 63.4 on DeepSWE v1.1, 48.8 on AutomationBench, 78.4 on Toolathlon Verified, and 55.3 on Humanity's Last Exam with tools, while Artificial Analysis independently puts it at 57 on its Intelligence Index. The review notes list pricing of $0.15 input, $0.50 output, and $0.03 cached input per million tokens, with a 50% launch promotion through September 9, 2026.

Z.ai released GLM-5.3-Flash on August 26, 2026, a 320B-parameter Mixture-of-Experts model that activates 18B parameters per token, reads text, images and video natively, and carries a 1M-token context window. It is the first natively multimodal model in the GLM-5 series and the first frontier open-weight release combining sparse and linear attention in a single stack, with weights published on Hugging Face under the MIT license. The architecture is a 45-layer stack, down from 92 in GLM-5.3, interleaving KDA linear-attention layers with NoPE sparse MLA layers, plus an IndexPool mechanism and Manifold-Constrained Hyper-Connections, routing each token through 8 of 288 experts. Native weights are FP8, roughly 306 GiB, and the model was pre-trained on a 30T-token multimodal corpus.

Z.ai introduced GLM-5.3-Flash as an open-weight multimodal model combining native support for several data modalities with an efficient Mixture-of-Experts structure that limits active computation, extending a lineage in which GLM-5.2 was the immediate predecessor. The open weights are released under the MIT license for direct download and modification from the Hugging Face repository. Prior anonymous community testing under the codename Ox Alpha on OpenCode and OpenRouter allowed early validation before the public announcement. The report positions the release at roughly one-tenth the cost of GLM-5.3 and one-fortieth that of Claude Opus 4.8, and notes that managed API endpoints on Z.ai's own platform provide a hosted alternative to self-hosting the weights.

Third-party benchmark aggregation places GLM-5.3-Flash at rank 31 on the composite LLM Stats Score, with category standings of 1 of 121 for chat, 23 of 231 for tool calling, 27 of 405 for reasoning, 34 of 352 for math, 37 of 242 for vision, and 59 of 304 for coding. It records a GDPval-AA Elo of 1773, ranked 2 under Artificial Analysis methodology, and a CharXiv-R score of 0.89 at rank 9. The same page shows measurable degradation on long conversations, with a score of 15.1 at turn 1 and an estimated 12.7 by turns 31+, and no data recorded for turns 11 through 30. Blended price is listed near $0.17 per million tokens against Claude Haiku 5.5 at $0.12, though the pricing basis differs from host-specific gateway rates.

Artificial Analysis lists GLM-5.3-Flash from Z AI as an open-weights model released in August 2026, scoring 28 on its Intelligence Index, above the 23 median for comparable open-weight models of similar size. It records 57 tokens per second output speed, a 92.39-second time to first answer token, 1M tokens of context, text and image input with text output, 320B total and 18B active parameters, and an MIT license for the weights. The independent measurement balances vendor claims by noting the model is notably slow per task at 39.5 minutes against a 14.2-minute median, and somewhat verbose, generating 124M tokens on the Intelligence Index versus a 99.3M median. Pricing is shown as $0.15 per million input tokens, $0.50 per million output tokens, with an 80% cache discount, averaging $0.58 per task.

Z.ai's official model card introduces GLM-5.3-Flash as the first natively multimodal model in the GLM-5 series, built on a newly trained base with 320B total parameters and 18B active parameters. It uses a hybrid architecture combining sparse and linear attention, plus Manifold-Constrained Hyper-Connections, trained on a 30T-token multimodal corpus, and outperforms GLM-5.2 at one-tenth the price while approaching Claude Opus 4.8 on coding and agentic benchmarks. Local deployment is documented for SGLang, vLLM, Transformers, KTransformers, Unsloth, and TokenSpeed, and the card explains that the reasoning effort parameter accepts low, high, and max levels, defaulting to max, while chat templates default clear thinking to false. Evaluation footnotes detail sampling settings and note that GPT-5.6-luna (medium) served as the judge model for HLE with tools.