GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
sinonchum
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
sinonchum
Model Tree
Qwen/Qwen3-4B-Instruct-2507
Input Modalities
Output Modalities
Supported Functionality
| Property | Value |
|---|---|
| Architecture | Qwen3-4B (GQA, RoPE, 8K) |
| Base Model | Qwen/Qwen3-4B-Instruct-2507 |
| Adapter Size | ~63 MB (rank=8) |
| Precision | bfloat16 |
| Context Length | 8,192 |
| License | Apache 2.0 |
Trained for three AI agent core functions:
| Parameter | Value |
|---|---|
| Method | LoRA (r=8, α=16, dropout=0.05) |
| Precision | bfloat16 |
| Dataset | Feiyue v11_8k (2,132 samples) |
| Max Seq Length | 8,192 tokens |
| Epochs | 3 |
| Learning Rate | 2e-4 (cosine, 10% warmup) |
| Effective Batch Size | 4 (1 × 4 grad accum) |
| GPU | NVIDIA L40S (48 GB) |
| Framework | Transformers 4.51 + PEFT 0.14 |
| Training Time | ~44 min |
| Final Train Loss | 0.6486 |
text
q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_projpython
from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizerimport torch base_model = AutoModelForCausalLM.from_pretrained( "Qwen/Qwen3-4B-Instruct-2507", torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True,)model = PeftModel.from_pretrained(base_model, "sinonchum/Qwen3-4B-Feiyue-v1")tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B-Instruct-2507") messages = [ {"role": "system", "content": "You are Feiyue, an AI agent."}, {"role": "user", "content": "Generate a PRD for a mobile app feature."},]text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer(text, return_tensors="pt").to(model.device)outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.7)print(tokenizer.decode(outputs[0], skip_special_tokens=True))| Variant | Type | Size | Repo |
|---|---|---|---|
| bf16 Merged | Full model | ~8 GB | bf16 |
| Q8_0 ⭐ | GGUF 8-bit | 4.0 GB | Q8_0 |
| Q4_K_M | GGUF 4-bit | 2.5 GB | Q4_K_M |
Evaluated on BFCL v4 exec categories using Modal L40S GPU. AST-based scoring against ground truth.
| Category | Accuracy | Correct | Total |
|---|---|---|---|
| exec_simple | 89.0% | 89 | 100 |
| exec_parallel | 88.0% | 44 | 50 |
| exec_multiple | 80.0% | 40 | 50 |
| exec_parallel_multiple | 82.5% | 33 | 40 |
| Overall | 85.8% | 206 | 240 |
bibtex
@misc{feiyue-v1-2026, author = {Simon Qin}, title = {Qwen3-4B-Feiyue-v1: LoRA Adapter for AI Agent Workflows}, year = {2026}, publisher = {Hugging Face}, howpublished = {\url{https://huggingface.co/sinonchum/Qwen3-4B-Feiyue-v1}},}from peft import PeftModelfrom transformers import AutoModelForCausalLM, AutoTokenizerimport torch base_model = AutoModelForCausalLM.from_pretrained( "Qwen/Qwen3-4B-Instruct-2507", torch_dtype=torch.bfloat16, device_map="auto", trust_remote_code=True,)model = PeftModel.from_pretrained(base_model, "sinonchum/Qwen3-4B-Feiyue-v1")tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-4B-Instruct-2507") messages = [ {"role": "system", "content": "You are Feiyue, an AI agent."}, {"role": "user", "content": "Generate a PRD for a mobile app feature."},]text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)inputs = tokenizer(text, return_tensors="pt").to(model.device)outputs = model.generate(**inputs, max_new_tokens=2048, temperature=0.7)print(tokenizer.decode(outputs[0], skip_special_tokens=True))@misc{feiyue-v1-2026, author = {Simon Qin}, title = {Qwen3-4B-Feiyue-v1: LoRA Adapter for AI Agent Workflows}, year = {2026}, publisher = {Hugging Face}, howpublished = {\url{https://huggingface.co/sinonchum/Qwen3-4B-Feiyue-v1}},}