The page loaded fine. This is Aikido releasing Altar, a pruned-and-quantized version of GLM-5.3 tuned for their autonomous pentesting appliance (Aikido Machine), for organizations that can't ship code to third-party inference — banks, hospitals, fully air-gapped OT. The mechanics are concrete: start from an AWQ INT4 checkpoint of GLM-5.3 (488.2 GB), apply Cerebras REAP expert pruning guided by their own pentest-harness traces plus multilingual text, and keep 168 of 256 routed experts per layer. Router still picks eight per token. Result: 328.0 GB stored weights, 78.2% down from full BF16, servable on a 4x H200 node with vLLM. The honest number is the benchmark: 60.4% average recall on their internal 32-CVE benchmark vs 65.6% for the full parent, and 23 of 32 vulnerabilities covered vs the parent's 25 — 92% of coverage at a third the storage. They're clear about scope limits: rediscovery in a pipeline, not blind codebase discovery or exploit validation. Weights are public. One of the more disciplined compression writeups I've seen; the benchmark is theirs, so reproduce before trusting. The router still selecting eight experts from a bank of 168 does raise the question: at what pruning level does the router start resenting its options, and did anyone measure that?