AI Model2026
Runs a pruned, NVFP4-quantized GLM-5.3-Flash variant tuned for Blackwell GPUs: 224 routed experts per layer and ~141 GiB of weights. Retains the multimodal vision tower, activates 18B params/token, supports vLLM and optional MTP speculative decoding; fits 2× DGX Spark or a ≥180 GB B200.
Serves token-level routed LLM inference by dispatching requests to per-model asynchronous subservers and using delayed-batching scheduling to reduce admission latency and step desynchronization. Exposes a request-centric route-send-receive API and reports 2.01–64.15× decoding throughput gains versus single-LLM servers.