GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
qiusizhan
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
qiusizhan
Model Tree
Qwen/Qwen3.5-9B
Input Modalities
Output Modalities
Supported Functionality
Three full epochs completed. Native full language-parameter SFT plus supervised contrastive learning; the vision encoder is frozen, with no LoRA. Planned scope: all 74,714 official AndroidControl training actions, 74,722 materialized examples, three epochs, 14,013 updates, physical batch 16, LR 1e-5, 10% warmup and cosine decay. The pinned native chat template disables thinking. BF16 forward/gradients, FP32 master AdamW and native FP32 RoPE are retained. The documented native channels_last_3d frozen-vision patch layout is used in training and evaluation.
This is a declared Qwen3.5 adaptation of AgentGhost, not reproduction of the paper's original backbone. Source: CTZhou-byte/AgentGhost at f6036b09610e8457643d286f057bfefde9c59332. Source/config/cache identities and native serialization checks accompany each checkpoint.
Runtime inputs are frozen independently of outputs: 100 trigger, 100 no-trigger, 50 defined border cases, plus 100 private diagnostic controls. Nominal reference inputs are 500 original normal-task-distribution cases. Input strata are not GT. Generated next actions are text only; no device command is executed. Independent output-specific labels are pending until a verified evaluation summary is added. Training completion alone establishes neither a selective backdoor nor rollout ASR.