GLM 4.6V Flash

Both are 9B lightweight versions of the GLM-4.6V series. GLM-4.6V-Flash is the free-tier variant with no formal SLA. GLM-4.6V-FlashX is a paid variant (¥0.05/M input tokens, ¥0.30/M output tokens as of June 2026) that offers SLA-backed reliability suitable for production workloads. Choose Flash for prototyping and cost-free pipelines; choose FlashX when you need guaranteed uptime.

In standard LLM tool use, images must be described in text before being passed as tool arguments — losing spatial, color, and layout information in the process. GLM-4.6V-Flash routes images, screenshots, and document pages directly as tool inputs (via URL references) and can process visual outputs returned by tools — such as charts, rendered web pages, or image search results — as part of the same reasoning chain. The API surface is identical to standard OpenAI tool calling; the multimodal capability is handled transparently by the model and the Z.ai / TheRouter layer.

The model's 128K token context window was trained natively (not extended post-hoc), which means it can sustain coherence across the entire window. The Z.ai team reports practical equivalents of ~150 pages of dense documents, 200 slide pages, or one hour of video in a single inference pass, because visual pages are encoded as tokens by the visual encoder. Maximum output is capped at 16,384 tokens per request.

The Z.ai team explicitly acknowledges the following in the model card: (1) pure-text QA performance is weaker than the visual capabilities, because this training cycle prioritised multimodal scenarios; (2) the model may occasionally over-think or repeat itself on complex prompts; (3) in some cases it restates the answer at the end of a response; (4) counting accuracy and identification of specific individuals remain imperfect. For free-tier, non-SLA workflows these are acceptable trade-offs. For production paths where any of these matter, validate outputs carefully or route to a paid model.