- Maker
- Z.ai (base GLM-5.3-Flash); uncensored build by dealignai
- Open weights
- Yes
- License
- MIT (base model); uncensored FP8 build republished by dealignai
- Context window
- 1,024K tokens
- Input images
- Supported — multiple images accepted per request
- Released
- August 26, 2026 (base model); uncensored FP8 weights August 2026; live on Venice since November 13, 2026
- Architecture
- Hybrid mixture-of-experts with sparse and linear attention; Manifold-Constrained Hyper-Connections (mHC)
- Parameters
- 320B total, 18B active per token
- Uncensoring method
- Abliteration — refusal behavior removed by editing the weights, no retraining
- Quantization on Venice
- FP8 (native block-FP8 build)
- Max output
- 128K tokens
- Capabilities
- Vision, Function calling, Reasoning, Web search, Code-optimized
- Privacy on Venice
- Private — zero retention
- Available on Venice since
- Nov 2026