Skip to main content
Back

Aikido Security's Altar-1 Shrinks a 1.5TB Security AI by 78%

Aikido Security released Altar-1, a 328 GB expert-pruned quantization of GLM-5.3 built for on-premises defensive cybersecurity workloads.

  • Aikido released Altar-1, a 328 GB open-weight security model pruned from GLM-5.3.
  • Uses REAP expert pruning (168 of 256 experts kept) plus AWQ INT4 quantization, no retraining.
  • Serves on 4x H200 GPUs via vLLM with 131k context length.
  • KL divergence of 0.506 nats vs full BF16 GLM-5.3 on sealed prompt panel.
  • Powers Aikido Machine, the company's on-prem autonomous pentesting appliance.
  • Inherits GLM-5.3 license: open weights with a large-operator commercial review clause.

Aikido Security has released Altar-1, an open-weight language model for security teams that need to keep source code, architecture documents, and vulnerability findings inside their own infrastructure. Aikido adapted GLM-5.3 rather than running a new pretraining cycle, reducing the checkpoint from 1.51 TB to 328 GB through quantization and expert pruning. The model also powers Aikido Machine, the company’s on-premises penetration-testing appliance.

How 1.51 TB Became 328 GB

GLM-5.3 is a 753-billion-parameter mixture-of-experts model. Each token activates 8 of 256 routed experts, using roughly 40 billion parameters per token. Its full BF16 checkpoint occupies 1,506.7 GB, which makes long-context serving at useful batch sizes impractical for many enterprise deployments.

Build Size Change from BF16
Full GLM-5.3 BF16 1,506.7 GB Baseline
AWQ INT4 base 488 GB About 68% smaller
Altar-1 after pruning 328 GB About 78% smaller

Aikido combined two compression methods to produce the final checkpoint:

  • AWQ INT4 quantization: The routed experts use 4-bit weights and 16-bit activations, based on the quantized GLM base. Attention, the shared expert, dense layers, and output head remain in BF16.
  • REAP expert pruning: Router-weighted Expert Activation Pruning scores each expert’s contribution and removes low-value experts without retraining. Altar-1 retains 168 of the original 256 routed experts in each backbone layer.

The pruning process preserves specialists by scoring each expert according to its largest share of routed work within any calibration domain. This approach reduces the chance that globally infrequent experts for code, uncommon languages, or structured output will be discarded. Aikido calibrated the selection with cybersecurity traces, code, tool calls, reasoning tasks, English text, and multilingual Wikipedia articles.

Routing still activates eight experts per token from the remaining pool of 168, so active parameter use stays near 40 billion. The smaller checkpoint reduces storage and memory requirements while preserving the model’s per-token compute profile.