GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
fyw1999
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
fyw1999
Model Tree
fyw1999/VisHarness-SFT-Qwen3-VL-8B-Thinking
Input Modalities
Output Modalities
Supported Functionality
The model is a multi-turn visual agent for GRES, ReasonSeg, and REC8K. It
coordinates PhraseToPoint, PhraseToBoxMask, PointToBoxMask,
SplitImageIntoPatches, SuperResolution, and MergeBoxMask.
Loading the weights alone does not execute these tools. Full agent inference
requires the VisHarness prompts, trajectory runner, and expert services:
VisHarness source code.
| Setting | Value |
|---|---|
| Algorithm | GRPO |
| Accepted prompt groups per update | 8 |
| Sampled trajectories per prompt | 8 |
| Reward-standard-deviation group filter | Enabled, threshold 0.2 |
| Policy-gradient weighting | Equal total trajectory weight |
| Loss aggregation | seq-mean-token-mean |
| Optimizer steps per global update | 2 |
| Learning rate | 1e-6, constant |
| Actor KL coefficient / estimator | 0.04 / low_var_kl+ |
| Sequence parallel size | 2 |
| Trajectory turn-count penalty | Disabled |
Step 550 had the highest mean task reward among the saved validations of
this run: 2.2264. Validation used a fixed subset of 300 examples, with
100 examples each from GRES, ReasonSeg, and REC8K, and one sampled trajectory
per example (temperature=0.6, top_p=0.95, top_k=20, up to 2048 output
tokens per turn).
| Task | Validation metric | Value |
|---|---|---|
| GRES | gIoU / cIoU | 61.41% / 67.98% |
| ReasonSeg | gIoU / cIoU | 74.55% / 67.97% |
| REC8K | MAE / RMSE | 7.78 / 16.36 |
These are sampled, subset-validation results, not full test-set results. The checkpoint is selected by mean task reward and is not the best checkpoint for every individual metric.
The repository is private. Authenticate with an account that has access, then download it with:
python
from huggingface_hub import snapshot_download snapshot_download( repo_id="fyw1999/VisHarness-GRPO-Qwen3-VL-8B-Thinking-step550", local_dir="./VisHarness-GRPO-pr6culeq-step550",)For BF16 inference, explicitly select BF16 in the model loader or serving configuration. Keep the VisHarness prompt and tool configuration aligned with the trained agent.
from huggingface_hub import snapshot_download snapshot_download( repo_id="fyw1999/VisHarness-GRPO-Qwen3-VL-8B-Thinking-step550", local_dir="./VisHarness-GRPO-pr6culeq-step550",)