AikidoSec/altar-1
AikidoSec/altar-1 is a 500.8 billion parameter mixture-of-experts (MoE) model, pruned from the GLM-5.3 architecture by 0xSero using Router-weighted Expert Activation Pruning (REAP). This model retains 168 of 256 experts per layer, with experts quantized to INT4 (W4A16) for efficient serving on Hopper GPUs. Calibrated on cybersecurity traces, coding, tool calling, reasoning, and multilingual data, Altar-1 is optimized for high-performance inference in specialized domains.
Loading preview...
Altar-1: A Pruned and Quantized GLM-5.3 MoE Model
Altar-1 is a 500.8 billion parameter model derived from the 753 billion parameter GLM-5.3 mixture-of-experts (MoE) architecture. Developed by 0xSero, it utilizes Router-weighted Expert Activation Pruning (REAP) to remove 34% of GLM-5.3's experts, resulting in 168 active experts per layer while maintaining the original 8 experts per token routing. This pruning method prioritizes retaining domain-specific experts, ensuring performance across diverse tasks.
Key Technical Details
- Architecture: Pruned GLM-5.3 MoE, with ~40B active parameters per token.
- Quantization: Experts are quantized to INT4 (W4A16) using AWQ, based on cyankiwi's GLM-5.3-AWQ-INT4. Attention and dense layers remain BF16.
- Calibration: Tuned on a diverse dataset including cybersecurity traces, coding, tool calling, reasoning, and multilingual Wikipedia articles.
- Fidelity: Demonstrates a KL divergence of 0.506 nats compared to the full BF16 GLM-5.3, indicating high fidelity despite significant pruning and quantization.
Optimized for Deployment
Altar-1 is specifically engineered for efficient serving on 4x H200 (Hopper) GPUs using vLLM, requiring 328 GB of weights. This configuration allows for a 128k-context KV cache at production batch sizes, with vLLM automatically selecting the Marlin MoE kernel. The model's design makes it suitable for high-throughput, specialized applications where resource efficiency and performance are critical.