Aikido releases Altar-1 open-weight GLM-5.3 security model

Aikido released Altar-1, an open-weight GLM-5.3 derivative compressed to 328 GB for on-prem vulnerability testing, with 60.4% internal recall versus 65.6%.

Aikido Security released Altar-1, an open-weight model for vulnerability discovery and penetration testing inside customer-controlled, on-premises or air-gapped infrastructure. Both reports describe it as a compressed derivative of Z.AI’s GLM-5.3; the Sept. 25 account specifies a 753-billion-parameter mixture-of-experts model, an AWQ INT4 starting checkpoint, and Cerebras REAP pruning that removes 88 of 256 routed experts per layer, leaving 168 experts and about 40 billion active parameters. Stored weights drop from 1,506.7 GB in BF16—rounded to 1.51 TB on Sept. 22—to 328 GB. On Aikido’s internal benchmark of 32 known vulnerabilities, which the earlier report says span 30 repositories, average recall was 60.4% versus 65.6% for the full-precision parent, and Altar-1 covered 23 cases. The Sept. 22 report alone says the parent rediscovered 25 flaws and that about 92% of coverage was retained; the later report does not give those two figures. Weights are on Hugging Face and can run in vLLM on four NVIDIA H200 GPUs, including the Aikido Machine appliance; the Sept. 25 report adds a 128k context and says the GLM-5.3 license permits commercial use subject to a large-revenue review.

• Aikido Security released Altar-1, an open-weight model for vulnerability discovery and penetration testing on customer-controlled, on-premises or air-gapped systems.
• It is derived from Z.AI’s 753-billion-parameter GLM-5.3 mixture-of-experts model and keeps about 40 billion active parameters.
• From an AWQ INT4 checkpoint, Cerebras REAP removes 88 of 256 routed experts per layer, leaving 168 experts.
• Stored weights fall from 1,506.7 GB in BF16 (about 1.51 TB in the Sept. 22 report) to 328 GB.
• On an internal benchmark of 32 known vulnerabilities across 30 repositories, recall was 60.4% versus 65.6% for the full-precision parent; Altar-1 covered 23 cases, and only the Sept. 22 report says the parent found 25 and that about 92% of…
• Weights are on Hugging Face and run in vLLM on four NVIDIA H200 GPUs, including the Aikido Machine appliance, with a 128k context under the GLM-5.3 license, which allows commercial use with a large-revenue review.