GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
visnia-ai
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
visnia-ai
Model Tree
Qwen/Qwen3.5-4B
Input Modalities
Output Modalities
Supported Functionality
| Model | Success rate | Relative output-token efficiency |
|---|---|---|
| Qwen/Qwen3.5-4B baseline | 22% | 1.0x |
| This model (SFT, FP8) | 62% | 4.6x |
The source data comprised 3.3k successful executor trajectories and auxiliary browser agent stages. Concise GPT-5.6 Luna reasoning summaries were retained where available.
| Setting | Value |
|---|---|
| Base model | Qwen/Qwen3.5-4B |
| Method | LoRA supervised fine-tuning |
| Epochs | 2 |
| LoRA rank / alpha | 64 / 128 |
| Effective global batch size | 8 |
The checkpoint was validated with vLLM 0.23.0 using the Qwen reasoning and tool call parsers:
bash
vllm serve "$MODEL_ID" \--served-model-name qwen3.5-4b-browser-agent-sft-fp8 \--tensor-parallel-size 1 \--max-model-len 262144 \--reasoning-parser qwen3 \--enable-auto-tool-choice \--tool-call-parser qwen3_coder
Use Qwen's recommended precise coding (WebDev) sampling profile:
json
{"temperature": 0.6,"top_p": 0.95,"top_k": 20,"min_p": 0.0,"presence_penalty": 0.0,"repetition_penalty": 1.0}