GLM-4.6V

GLM-4.6V, published in December 2025, is the clearest example in Z.ai's catalogue of a generational jump that costs nothing in hardware. The safetensors index on Hugging Face reports 107,710,933,120 parameters — byte for byte the figure published for GLM-4.5V four months earlier. The network was not enlarged; it was retrained. What changed is everything around the weights. The context window doubled from 64,000 to 128,000 tokens, so a full document set or a long screen recording fits in one request. Native function calling arrived for the first time in a GLM vision model, which is the feature the company itself flags as the point of the release: it closes the gap between a model that sees a screen and an agent that acts on it. And the price halved, from $0.60 and $1.80 per million tokens to $0.30 and $0.90. Z.ai sells three tiers. The standard model is the one described above; GLM-4.6V-FlashX is a lightweight, faster version at $0.04 input and $0.40 output; GLM-4.6V-Flash is free of charge. All three take video, images, text and files, and all three carry the 128,000-token window — unusual, since vendors normally shorten the context on their free tiers. The weights are on Hugging Face under an MIT licence. Downloads stood near 8,400 at the time of writing, an order of magnitude behind its predecessor — the usual lag for a model that has been out for a shorter time and whose predecessor is already deployed everywhere.

▸Videos

No videos yet.

▸Reviews