Aikido Security Releases Altar-1: An Open-Weight Security Model Pruned From GLM-5.3 to 328 GB
Aikido released Altar-1, an open-weight security model pruned from GLM-5.3 to 328 GB.
Aikido Security published Altar-1, a compressed derivative of Z.AI’s 753-billion-parameter GLM-5.3 mixture-of-experts model, for on-prem and air-gapped security use. It starts from an AWQ INT4 checkpoint and uses Cerebras REAP to remove 88 of 256 routed experts per layer, cutting weights from 1,506.7 GB in BF16 to 328 GB while keeping about 40 billion active parameters. On Aikido’s 32-vulnerability internal benchmark, average recall fell from 65.6% to 60.4%, still covering 23 of 32 cases. Weights are on Hugging Face and run in vLLM on four NVIDIA H200 GPUs under the GLM-5.3 license.
- Altar-1 keeps 168 of 256 experts and about 40 billion active parameters.
- Stored weights fall from 1,506.7 GB in BF16 to 328 GB.
- CVE-benchmark recall is 60.4 percent versus 65.6 percent for BF16.
- Deployment targets four H200 GPUs with vLLM and 128k context.
- The GLM-5.3 license allows commercial use with a large-revenue review.
- Products
- vLLMAikido Machine
- Organizations
- Aikido SecurityZ.AIHugging FaceCerebras
From the original article
4 min read at marktechpost.com
Aikido Security has released Altar-1, its first open-weight security model. It is a compressed version of Z. AI’s GLM-5.
By Asif Razzaq · marktechpost.com
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.