GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
zyhang1998
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
zyhang1998
Model Tree
Qwen/Qwen3.5-397B-A17B
Input Modalities
Output Modalities
Supported Functionality
bash
vllm serve zyhang1998/qwen3.5-397b-a17b-mlsbench-syn-rl-rewardfix-step104 \--tensor-parallel-size 8 \--dtype bfloat16 \--reasoning-parser qwen3 \--speculative-config '{"method":"mtp","num_speculative_tokens":2}'
Qwen/Qwen3.5-397B-A17BThe checkpoint and validation artifacts completed successfully, but the
post-run policy-consistency gate did not pass. At step 104,
training/rollout_probs_diff_mean was 0.96975,
training/rollout_probs_diff_max was 1.0, and the reported rollout/training
KL was 60.1778. Treat this as an experimental research checkpoint and verify
behavior on the target workload before deployment.
For the base architecture, intended use, limitations, and license terms, see the Qwen3.5-397B-A17B model card.