GLM-4.6V: nominal 106B total / 12B active MoE (107.711B stored including vision/projector); Flash: nominal 9B dense (10.293B including vision/projector)
128K (131,072 tokens)
MIT (publisher-declared weights); AIMv2 encoder upstream rights require review
Same-launch hybrid-reasoning vision-language family: GLM-4.6V and its official FP8 package, plus dense GLM-4.6V-Flash. The updated GLM-V report identifies GLM-4.5-Air and GLM-4-9B-0414 language backbones respectively, with an AIMv2-Huge-initialized vision encoder and visual-to-language adapter. The main decoder has 46 layers, 128 routed experts with eight selected per token and one shared expert, and 96 query / eight KV heads; Flash has 40 dense layers and 32 query / two KV heads. Both accept text, images and video and generate text, with thinking/non-thinking control. Native multimodal function calling is implemented through XML tool calls and image/video-bearing tool responses; URL-referenced visuals enable tool-assisted search, cropping and interleaved illustrated reports, not native image synthesis or built-in execution of arbitrary tools. Documents are understood as rendered page images; file parsing, retrieval and rendering still require the surrounding application. The launch advertises screenshot-to-HTML/CSS/JS replication, visual edits, long-document analysis and approximately 150 document pages, 200 slides or a one-hour video in one pass, dependent on visual tokenization. Training includes multimodal pre-training, long-context continual pre-training, long-CoT supervised fine-tuning and multimodal reinforcement learning with curriculum sampling; the launch additionally describes billion-scale perception/world-knowledge data, synthetic agentic data and tool-invocation rewards. The January 2026 report update explicitly describes a 131,072-token extension and SFT for GLM-4.6V; exact checkpoint-specific training budgets and Flash agentic-stage details are not fully disclosed. Launch evaluation covers over 20 multimodal benchmarks; the retrospective report gives main-model thinking scores of 76.0 MMMU, 85.2 MathVista and 54.9 MMLongBench-Doc, with both gains and regressions versus GLM-4.5V rather than universal superiority on 42 tasks. Agent and coding results depend on external scaffolds and judges; launch-recommended generation uses temperature 0.8, top_p 0.6, top_k 2, repetition penalty 1.1 and a 16K output budget, not a separate native context limit. The publisher notes weaker pure-text QA, overthinking/repetition and counting or identity-recognition errors.