GLM-4.6V: Z.ai's GLM-4.6 Vision Model, 128K Context

What GLM-4.6V is

GLM-4.6V is the vision-language member of Z.ai's GLM-4.6 generation: MIT-licensed, self-hostable, with a 128K token context window, at roughly $0.30 per million input tokens and $0.90 output through third-party providers. The V suffix across this family marks the multimodal releases, and Z.ai have published one for most generations — GLM-4.5V precedes it and GLM-5V Turbo follows, the latter being the one exception in an otherwise MIT-licensed line, since it is proprietary.

LMArena

On LMArena's text leaderboard GLM 4.6V has 1378 Elo (rank 194); by category, 1416 Elo in Coding (rank 201), 1395 Elo in Expert (rank 191), 1350 Elo in Creative Writing (rank 182) and 1367 Elo in Instruction Following (rank 193). LMArena's text board presents written prompts only; it does not measure visual performance.

Pricing

At roughly $0.30 per million input tokens and $0.90 output through third-party providers, GLM-4.6V costs about five times GLM-4.7-Flash and roughly twice GLM-4.5-Air. Qwen3-VL-235B-A22B-Instruct runs around $0.21 per million input and has 235 billion parameters to self-host.

Two sizes and native function calling

The GLM-4.6V series has two versions: GLM-4.6V, a 106-billion-parameter foundation model for cloud and high-performance use, and GLM-4.6V-Flash, a 9-billion-parameter model for local deployment and low-latency applications. Z.ai state the context window was scaled to 128K tokens in training, and that GLM-4.6V is the first in the line with native function calling: images, screenshots and document pages can be passed directly as tool inputs without text conversion, and visual tool outputs such as charts or rendered pages are interpreted in the reasoning chain.

Capabilities and stated limitations

The card describes interleaved image-and-text content generation from multimodal context, with search and retrieval calls during generation; long-document understanding of up to 128K tokens, reading formatted pages as images with text, layout, charts, tables and figures together; and reconstruction of HTML and CSS from UI screenshots with natural-language edits. Z.ai list limitations: pure text question answering has significant room for improvement, the model may overthink or repeat itself on complex prompts, it may restate the answer at the end, and counting and identifying specific individuals remain weak. To reproduce their results Z.ai recommend top_p 0.6, top_k 2, temperature 0.8, a repetition penalty of 1.1 and up to 16K generated tokens, with vLLM as the main backend and SGLang for video. The model runs on Transformers 5.0.0rc0 or later, vLLM 0.12.0 or later and SGLang 0.5.6.post1 or later.

On the Z.ai API

Z.ai's API documentation lists video, image, text and file input with text output, and states that the 128K context covers roughly 150 pages of complex documents, 200 slide pages or an hour of video in a single pass. Listed capabilities include visual web search as an end-to-end multimodal search-and-analysis workflow, front-end replication of designs to HTML, CSS and JavaScript, and object detection with bounding-box output. Z.ai evaluated it on more than 20 multimodal benchmarks including MMBench, MathVista and OCRBench. The API also offers GLM-4.6V-FlashX as a lighter paid option and GLM-4.6V-Flash free of charge, both with the same 128K context.

Official Resources