GLM-5.3 Flash: fast, multimodal and cheap

GLM-5.3 Flash was listed on OpenRouter on 26 August 2026. Unlike GLM-5.3, it starts from a newly trained base model whose architecture and training recipe were redesigned for capability and efficiency. It has 320 billion parameters in total, of which only 18 billion are active per token.

For the first time in the GLM series, Z.ai combines sparse and linear attention in a hybrid design, which sharply reduces long-context serving cost while keeping precise long-context behaviour. Z.ai says this reduces attention compute and KV cache by 3.01× and 4.44× compared with GLM-5.3. The model also uses Manifold-Constrained Hyper-Connections (mHC) and was pre-trained on a 30-trillion-token multimodal corpus.

On OpenRouter, GLM-5.3 Flash accepts text, images and video, with a context window of 1,048,576 tokens and a maximum output of 943,718 tokens. It costs $0.0052 per million input tokens and $0.1985 per million output tokens — the lowest prices in the GLM-5 family by a wide margin. On Z.ai's platform the model code is glm-5.3-flash, with a 1M context, up to 128K output tokens and image input through image_url.

Z.ai claims GLM-5.3 Flash outperforms GLM-5.2 across benchmarks and real-world workloads at one-tenth of the price, and approaches Claude Opus 4.8 on coding and agentic benchmarks. The weights are on Hugging Face as zai-org/GLM-5.3-Flash under the MIT licence.

GLM-5.3 FlashX, listed on 18 September 2026, is the high-speed variant of GLM-5.3 Flash. It is built on the same hybrid attention architecture and delivers inference speeds of up to 200 tokens per second, according to its description. It accepts text, images and video with a 1M-token context but returns at most about 131K tokens per request.

FlashX is priced at $0.37 per million input tokens and $1.25 per million output tokens on OpenRouter and has no public weights. Choose it when latency matters more than cost; for most other work, plain GLM-5.3 Flash is far cheaper.