GLM 5.3 Flash vs Qwen3.8 Flash

GLM-5.3-Flash is the first natively multimodal model in Z.ai's GLM-5 series, a mixture-of-experts transformer with roughly 320 billion total parameters and 18 billion activated per token. It routes each token through 8 of 288 experts across 45 language layers that interleave KDA linear attention with sparse multi-head latent attention, and pairs them with a 24-layer vision encoder that handles image and video input. The checkpoint declares a maximum context length of 1,048,576 tokens, ships in native FP8, and includes a multi-token prediction draft layer for speculative decoding. Z.ai reports that the hybrid attention design reduces attention computation by 3.01x and KV cache size by 4.44x relative to GLM-5.3.

The model starts from a newly trained base built on a 30 trillion token multimodal pre-training corpus and adopts Manifold-Constrained Hyper-Connections to improve scaling efficiency. Vision is integrated into the coding and agent loop, so the model can inspect interfaces, rendered output, and images while operating across code, browsers, and graphical user interfaces. Z.ai reports scores of 84.3 on Terminal-Bench 2.1, 63.4 on DeepSWE 1.1, 55.3 on Humanity's Last Exam with tools, and 48.8 on AutomationBench, and the model exposes low, high, and max thinking modes.

Qwen3.8-Flash is a multimodal mixture-of-experts model from the Qwen team at Alibaba, and the production counterpart of the open-weight Qwen3.8-Flash-Next preview that introduces the architecture intended for the Qwen4 family. The main model carries 125 billion parameters alongside a separate 51 billion parameter N-gram embedding table, while activating roughly 6 billion parameters per token. It accepts interleaved image and text input and returns text, handling 262,144 tokens of context natively with extension to 1,000,000 tokens using YaRN. The production configuration runs with the 1M context window by default and adds built-in tool support.

Four architectural changes separate it from earlier Qwen releases: hybrid attention that pairs Gated DeltaNet for history compression with Qwen Sparse Attention, which uses a lightweight indexer to select micro-blocks of context; a Gated Residual scheme; N-gram embeddings; and training with the Muon optimizer, refined around orthogonalization accuracy and the division of parameters between Muon and AdamW. Qwen reports training cost around one ninth that of Qwen3.7-Plus, with QSA attention kernels measured up to 7.6 times faster in prefill and 4.9 times faster in decode at 1M-token context. Reported scores include 58.7 on DeepSWE 1.1, 62.5 on SWE-bench Pro, 84.5 on AndroidWorld and 95.7 on MathVision.