DownloadAltar-1 is Aikido Security’s open-weight cybersecurity-focused language model, created by pruning GLM-5.3 for more practical self-hosted deployment while retaining its core reasoning capabilities. Starting from GLM-5.3’s 753B-parameter Mixture-of-Experts architecture, Altar removes 34% of routed experts using REAP, leaving 168 of 256 experts per layer and approximately 504B parameters. Eight experts remain active per token, preserving roughly 40B active parameters during inference. The model was calibrated using cybersecurity traces, coding, tool calling, reasoning, English, and multilingual Wikipedia data to preserve domain specialists during pruning. Its routed experts use INT4 W4A16 AWQ quantization, while attention, shared experts, dense layers, and the output head remain BF16. The resulting weights occupy about 328 GB and are designed for production deployment on four NVIDIA H200 GPUs through vLLM, including 128K-context workloads.
Features
- Approximately 504B parameters after expert pruning
- 168 routed experts retained from GLM-5.3's original 256
- Eight experts active per token with roughly 40B active parameters
- REAP expert pruning without model retraining
- Calibrated on cybersecurity, coding, tool use, and reasoning data
- INT4 W4A16 AWQ quantization for routed experts
- 328 GB model designed for four NVIDIA H200 GPUs
- Supports 128K-context production serving through vLLM
Project Samples
