GLM-5.3 Flash (flexai)

A practitioner breakdown of GLM-5.3 Flash, released by Z.ai on 26 August 2026, describes it as the first GLM-5 series model trained on text and images from the start, structured as a mixture of experts with 320B total parameters and 18B active per token. It also breaks down how the model mixes two cheaper attention mechanisms to reduce the cost of serving long prompts. Z.ai's own reported numbers show GLM-5.3 Flash ahead of GLM-5.2 on Terminal-Bench 2.1 (84.3 vs 81.0), DeepSWE v1.1 (63.4 vs 46.2), AutomationBench v1.0.6 (48.8 vs 26.2), and Humanity's Last Exam with tools (55.3 vs 54.7), though the author notes these were not independently re-run. The spec sheet lists a 1M-token context window, 131,072-token max output, reasoning on by default, tool calling support, and an MIT license.

Z.ai's own technical post explains that GLM-5.3-Flash was trained from a new base model with a redesigned recipe, mixing linear attention for local dependencies with sparse attention that uses a lightweight indexer to pull relevant information from global context, and adding mHC for scaling efficiency on a 30T-token multimodal corpus. Relative to the GLM-4.5 series, active parameters drop from 32B to 18B and layers from 92 to 45. Against GLM-5.3, Z.ai reports roughly a 3.0x reduction in attention compute and a 4.4x smaller KV cache, with IndexPool pooling four cached key vectors into one to cut indexer latency and memory. Published scores include Terminal Bench 2.1 at 84.3 and DeepSWE v1.1 at 63.4, ahead of GLM-5.2's 81.0 and 46.2.

An independent technical deep dive describes GLM-5.3-Flash as a 320B mixture-of-experts model activating 18B per token, running natively in FP8 with a 1,048,576-token context window, and the first natively multimodal model in the GLM-5 series. It notes the model was tested anonymously as "Ox Alpha" on OpenRouter for roughly twelve days before launch. The post also records that Z.ai's full GLM-5.3 flagship was announced on August 14, 2026 with 744B open weights promised but no GLM-5.3 repository appearing on Z.ai's Hugging Face organization at launch time. It lists MIT-licensed weights on Hugging Face, a public vLLM recipe, and API pricing of $0.15 input and $0.50 output per million tokens.

Vals AI lists GLM 5.3 Flash as an open-weight model from zAI released August 26, 2026, with a one-million-token context window, 128k maximum output tokens, text and image input support, and token costs of $0.15 input and $0.50 output. It records a Vals Index accuracy of 50.86% plus or minus 1.25, placing it 26th of 32 ranked models. Independent sub-scores include a Vals Legal Index of 71.05%, SAGE at 89.25%, Code Migration v1.1 at 51.65%, CyberBench v1.1 at 67.86%, and Finance Agent v2.1 at 60.59%. The page also lists a cost per test of $2.049 on the Vals Index and a measured latency of 63 minutes 51 seconds.

Z.ai introduced GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series, with 320B total parameters and 18B active. It starts from a newly trained base model rather than a post-train of an earlier checkpoint, and the company says it outperforms GLM-5.2 across benchmarks and real-world workloads at roughly one-tenth the price. The architecture combines sparse and linear attention for cheaper long-context serving, and adds Manifold-Constrained Hyper-Connections (mHC) for better scaling efficiency, with training on a 30T-token multimodal pre-training corpus. Z.ai reports an Artificial Analysis Intelligence Index v4.1.1 score of 57 at $0.045 per task, and notes the model was tested anonymously as ox-alpha on OpenCode and OpenRouter.