Aikido Security has introduced Altar-1, an open-weight security model derived from the GLM-5.3 architecture. The model is compressed to 328 GB and is designed to run on a single node equipped with 4x NVIDIA H200 GPUs using vLLM.
The compression process involved two stages, including calibration using pentesting traces and various text datasets. While the original GLM-5.3 model contains 753B parameters, Altar-1 maintains 40B active parameters by selecting 8 out of 168 experts per token.
Performance testing on an internal CVE benchmark showed that Altar-1 retains 92% of the vulnerabilities covered by the parent model. The company reports that the model successfully identified a critical-severity vulnerability during a real-world client pentest.
Altar-1 requires Hopper-series GPUs, specifically H100 or H200 models. Due to its 328 GB size, it requires more than the 320 GB of memory available on a standard 4x H100 80 GB node, leaving space for a 128k-context KV cache on H200 hardware.
The model is released under the GLM-5.3 license, which allows for commercial use and modification. Operators with annual revenue exceeding $10 billion must undergo a security review. Aikido plans to explore lower-bit formats and fine-tuning for security workflows.
