GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
Gyubeum
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
Gyubeum
Model Tree
Qwen/Qwen3-VL-8B-Instruct
Input Modalities
Output Modalities
Supported Functionality
Gyubeum/UI_Genie_128k, sft/ split: 127,976 samples from 63,988 pairs.text
[system] You are a helpful assistant. + mobile_use tool spec (with screen resolution)[user] The user query: <goal> Task progress (You have done the following operation on the current device): Step1: <action> <image> ... StepN: <action>; <image>[assistant] <tool_call>{"name": "mobile_use", "arguments": {...}}</tool_call>[user] Was the agent's action above correct given the current screen state? Answer with exactly one token: <|+|> for correct, <|-|> for wrong.[assistant] ← model generates <|+|> or <|-|>A process reward model for evaluating GUI agent actions on mobile UIs:
For a continuous reward, see the Bradley-Terry models above.
The checkpoint loads like a standard Qwen3-VL model in transformers or vLLM; no adapter or extra head is needed.
python
import torchfrom PIL import Imagefrom transformers import AutoProcessor, Qwen3VLForConditionalGeneration MODEL_PATH = "Gyubeum/Qwen3-VL-8B-Instruct-UI-Genie-SFT-128K"# SYSTEM_PROMPT, goal, history and tool_call_response come from the agent step being judged.JUDGE_PROMPT = ( "Was the agent's action above correct given the current screen state? " "Answer with exactly one token: <|+|> for correct, <|-|> for wrong.") model = Qwen3VLForConditionalGeneration.from_pretrained( MODEL_PATH, torch_dtype=torch.bfloat16, device_map="cuda",).eval()processor = AutoProcessor.from_pretrained(MODEL_PATH, max_pixels=1_048_576)tok = processor.tokenizerpos_id = tok.convert_tokens_to_ids("<|+|>")neg_id = tok.convert_tokens_to_ids("<|-|>") messages = [ {"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": [ {"type": "text", "text": f"The user query: {goal}\nTask progress (...): {history}"}, {"type": "image", "image": Image.open("screenshot.png").convert("RGB")}, ]}, {"role": "assistant", "content": tool_call_response}, {"role": "user", "content": JUDGE_PROMPT},]inputs = processor.apply_chat_template( messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt",).to(model.device) with torch.no_grad(): logits = model(**inputs).logits[0, -1]top = logits.argmax().item()hard_score = 1.0 if top == pos_id else 0.0 if top == neg_id else 0.5soft_score = torch.softmax(logits[[pos_id, neg_id]].float(), dim=-1)[0].item()| Field | Value |
|---|---|
| Base model | Qwen/Qwen3-VL-8B-Instruct, with <|+|> / <|-|> added as special tokens |
| Training method | SFT, cross-entropy on the judge token (LoRA, merged) |
| Training data | 127,976 single-label samples from 63,988 preference pairs (Gyubeum/UI_Genie_128k sft/) |
| Epochs / steps | 1 epoch, 1,000 steps |
| Batch | 128 samples per step (1 per GPU × 16 grad-accum × 8 GPUs) |
| Trainable | LoRA r=64, α=32, dropout 0.05 on all LLM linear layers + full lm_head (796.9M params); vision tower and aligner frozen |
| Optimizer | AdamW, lr 1e-4, cosine schedule, 5% warmup, weight decay 0.1, grad clip 1.0 |
| Max length | 8,192 tokens (left truncation), max_pixels 1,048,576 |
| Precision / parallelism | bfloat16, DeepSpeed ZeRO-3, 8 GPUs |
| Output | Discrete preference token: <|+|> / <|-|> |
| Scoring | Greedy-decode 1 token → 1.0 / 0.0 / 0.5, or P(<|+|>) as a soft score |
Training code: train_rm/rm_main_scripts/sft/train_sft_128k.sh in AndroidFlux.
bibtex
@misc{qwen3technicalreport, title={Qwen3 Technical Report}, author={Qwen Team}, year={2025}, eprint={2505.09388}, archivePrefix={arXiv}, primaryClass={cs.CL},}[system] You are a helpful assistant. + mobile_use tool spec (with screen resolution)[user] The user query: <goal> Task progress (You have done the following operation on the current device): Step1: <action> <image> ... StepN: <action>; <image>[assistant] <tool_call>{"name": "mobile_use", "arguments": {...}}</tool_call>[user] Was the agent's action above correct given the current screen state? Answer with exactly one token: <|+|> for correct, <|-|> for wrong.[assistant] ← model generates <|+|> or <|-|>import torchfrom PIL import Imagefrom transformers import AutoProcessor, Qwen3VLForConditionalGeneration MODEL_PATH = "Gyubeum/Qwen3-VL-8B-Instruct-UI-Genie-SFT-128K"# SYSTEM_PROMPT, goal, history and tool_call_response come from the agent step being judged.JUDGE_PROMPT = ( "Was the agent's action above correct given the current screen state? " "Answer with exactly one token: <|+|> for correct, <|-|> for wrong.") model = Qwen3VLForConditionalGeneration.from_pretrained( MODEL_PATH, torch_dtype=torch.bfloat16, device_map="cuda",).eval()processor = AutoProcessor.from_pretrained(MODEL_PATH, max_pixels=1_048_576)tok = processor.tokenizerpos_id = tok.convert_tokens_to_ids("<|+|>")neg_id = tok.convert_tokens_to_ids("<|-|>") messages = [ {"role": "system", "content": SYSTEM_PROMPT}, {"role": "user", "content": [ {"type": "text", "text": f"The user query: {goal}\nTask progress (...): {history}"}, {"type": "image", "image": Image.open("screenshot.png").convert("RGB")}, ]}, {"role": "assistant", "content": tool_call_response}, {"role": "user", "content": JUDGE_PROMPT},]inputs = processor.apply_chat_template( messages, tokenize=True, add_generation_prompt=True, return_dict=True, return_tensors="pt",).to(model.device) with torch.no_grad(): logits = model(**inputs).logits[0, -1]top = logits.argmax().item()hard_score = 1.0 if top == pos_id else 0.0 if top == neg_id else 0.5soft_score = torch.softmax(logits[[pos_id, neg_id]].float(), dim=-1)[0].item()@misc{qwen3technicalreport, title={Qwen3 Technical Report}, author={Qwen Team}, year={2025}, eprint={2505.09388}, archivePrefix={arXiv}, primaryClass={cs.CL},}