Aikido Security has released Altar-1, an open‑weight security model derived from Z.AI’s GLM‑5.3 mixture‑of‑experts model. The core problem for many organizations is that frontier models must run on external infrastructure, forcing source code, architecture diagrams, and unremediated findings to leave the network. This violates data‑residency rules for banks, OT operators, and any air‑gapped environment. Open‑weight models solve the residency issue but introduce a deployment gap: the full MoE model stores all experts even when only a few are used per token, and long‑running security agents need GPU memory for KV cache that competes with the model weights.
Aikido built Altar-1 in two steps. First, they started from an AWQ INT4 checkpoint that stores routed expert weights in 4 bits while keeping attention, shared experts, dense layers, and the head in BF16 (W4A16). Second, they applied Cerebras REAP (Router‑weighted Expert Activation Pruning) which scores experts by router weight and output magnitude, not just selection frequency. This pruned 88 of the 256 routed experts per layer, leaving 168 experts and reducing stored weights from 488.2 GB to 328.0 GB—32.8 % smaller than the AWQ parent and 78.2 % smaller than the original BF16 model. Routing remains unchanged: each token still activates eight experts, yielding roughly 40 B active parameters.
The resulting model fits on a single node of four NVIDIA H200 GPUs (total 564 GB HBM). With Altar-1’s 328 GB weight footprint, about 236 GB remains for KV cache and runtime overhead, enough for long‑running pentesting agents. Deployment is straightforward: use vLLM with tensor‑parallel size four, trust remote code, and a max model length of 131072 tokens. No retraining is required, and the model can be served entirely within the customer’s controlled infrastructure, keeping all security context inside the network.
#AI #Product #CyberSecurity #LLM #OnPrem #DevSecOps