GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
laion
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Run this model inference with full control and performance in your environment.
Model Details
Model Provider
laion
Model Tree
Qwen/Qwen3-Coder-30B-A3B-Instruct
Input Modalities
Output Modalities
Supported Functionality
global_step_42 — highest trailing-5 reward EMA (alpha=1/3) over the stitched two-generation W&B
curve (steps 1-92, no gaps): EMA 0.2268; step reward 0.2520; pass@8 0.375; entropy 0.0237. This
matches the owner's "best single reads 42-43" note. All 31 banked checkpoints (steps 3-90) were
retained, so the true EMA-best was available (unlike the Jupiter KL arms). Exported on-Iris via the
checkpoint-export job at the saved policy geometry (2x8 H100, Megatron profile).
The arm cleared its early gate and improved to ~0.25 reward, then collapsed in epoch 2 from ~step 77: reward 0.02-0.05 for steps 77-88, entropy at 0.06x step 1, grad norm decaying — the POLICY stop condition. Verdict: an entropy bonus of 0.003 does NOT hold exploration on this schedule. Not an admissible horizon result; preserved as the entropy-lever negative control. Stopped by owner decision 2026-08-18 (job killed at logged step 92; banked 90).
See training_logs/ for metrics.csv (stitched per-step curve + trailing-5 EMA), report.md,
reward_plot.png, and the launch rl_config.yaml. W&B: dogml/OpenThoughts-Agent runs zgyhu8oc +
vne6jh9g (r2 generations).
open-athena/tt-x16-entropy0p003 — 1/16 systematic subsample (every 16th trial of 51,701, uniform coverage of steps 1-92). The full set (~248 GB) and 1/8 (~31 GB) exceeded the sync host's free disk (28 Gi); smallest-stride-that-fits per the owner-precedented deviation rule (x10/x15 used 1/8 under more headroom).