Altar-1

Altar-1 is Aikido Security's first open-weight model, a compressed version of Z.AI's GLM-5.3 for defensive security work inside infrastructure the customer controls, including air-gapped networks. Aikido announced it on 21 September 2026. It powers Aikido Machine, the company's on-premises autonomous penetration-testing appliance, and Aikido says the approach is being extended to Aikido Attack, Code Security Audit and Deep PR Review.

Aikido started from a community AWQ INT4 quantisation of GLM-5.3 and removed experts using Cerebras's REAP pruning method, calibrated on traces from its own pentesting harness plus coding, tool-calling, reasoning and multilingual text; the company says no customer data was used. Each expert layer keeps 168 of the original 256 routed experts, and eight are still selected per token. The model card describes about 504 billion parameters, roughly 40 billion active, with routed experts stored at 4 bits (W4A16). Stored weights fall from 1.51 TB for the full-precision parent to 328 GB. Aikido says it can be served on one node of four NVIDIA H200 GPUs with vLLM; its example command sets a 131,072-token context.

In its own evaluation on 32 known vulnerabilities across 30 repositories, Aikido reports average recall of 60.4% per run, against 61.5% for the quantised parent and 65.6% for full-precision GLM-5.3, and says Altar rediscovered 23 of the 32 at least once. The company notes that this measures targeted rediscovery, not blind discovery or exploit validation.

The weights are free to download under the GLM-5.3 licence, which allows commercial use and modification but requires a company that runs a Model-as-a-Service business and has revenue above $10 billion over any 12 months to pass a Z.AI security review before commercial use. Reuters reported the launch on 21 September.