GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Xnhyacinth
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Xnhyacinth
Model Tree
Qwen/Qwen3.5-4B
Input Modalities
Output Modalities
Supported Functionality
Full-parameter SFT of Qwen3.5-4B (instruct) on the ACC 128k related-work baseline. Companion to the 4B-Base run; same recipe length and step count.
| Field | Value |
|---|---|
| Checkpoint | checkpoint-680 (data/sft/swift_ext_acc/v6-20260825-153450) |
| Recipe | GBS 16, 8 GPU SP=4 DP=2, lr 1e-5 cosine, max_length 133120, 680 steps |
| Train loss | 0.365 |
| Eval loss | 0.375 |
| Data | ACC ShareGPT → acc.jsonl (see Xnhyacinth/longworld-128k-sft-baselines) |
This is the last step-680 weights, not an intermediate save. Optimizer/RNG
states are omitted. production_eligible is unchanged; this is a related-work
baseline, not a LongWorld product checkpoint.