Aikido Security Releases Open-Weight Model Pruned From GLM-5.3

Aikido Security has released Altar-1, its inaugural open-weight security model, on March 14, 2024. This model is a compressed iteration of Z.AI's GLM-5.3, engineered for deployment within customer-controlled infrastructure. Altar-1 serves as the foundational technology for Aikido Machine, the company's autonomous pentesting appliance, which is capable of operating on on-premise and air-gapped networks. The model's weights are publicly available on Hugging Face and can be executed using vLLM on a single node equipped with four NVIDIA H200 GPUs, indicating its deployability.

The development addresses a critical challenge in the security sector: the inability for sensitive security context to remain within a company's network. Traditional closed-frontier models require data to be processed on external infrastructure, potentially exposing source code, architecture documentation, and unremediated security findings. This poses a significant issue for entities like banks with strict data-residency mandates or Operational Technology (OT) operators lacking internet connectivity. While open-weight models resolve data residency concerns, they often present deployment hurdles, particularly with Mixture-of-Experts (MoE) architectures where all experts must be stored, even if only a subset is actively used for a given task. Furthermore, security agents require substantial long-running context, and the associated KV cache competes with model weights for limited GPU memory.

Altar-1 was constructed through a two-step compression process applied to GLM-5.3, a 753 billion parameter MoE model where each token routes to 8 out of 256 experts per layer, resulting in approximately 40 billion active parameters. The first step involved quantization, starting with the cyankiwi GLM-5.3-AWQ-INT4 checkpoint. This AWQ method stores routed expert weights in 4-bit precision, while activations are maintained at 16-bit. Shared components like attention, dense layers, and the output head are preserved in BF16 precision. The second step utilized Cerebras REAP (Router-weighted Expert Activation Pruning) for expert pruning. REAP evaluates experts based on router weights and output magnitude, not solely on their selection frequency. This process resulted in Altar-1 retaining 168 out of the original 256 routed experts per layer, effectively removing 88 experts, which constitutes a 34.4% reduction. Importantly, this pruning was achieved without any retraining of the model. The calibration for this process incorporated traces from Aikido's pentesting harness, alongside data from coding tasks, tool calling, reasoning exercises, and multilingual Wikipedia text. Aikido has stated that no customer data was utilized in this calibration phase. Each expert was scored based on its largest contribution to any single domain's routed workload, a methodology designed to safeguard the model's performance across diverse security-related tasks.