GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
hungta
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
hungta
Model Tree
Qwen/Qwen3-4B-Instruct-2507
Input Modalities
Output Modalities
Supported Functionality
LoRA adapter (r=16, alpha=32, 1 epoch, lr 1e-4, seed 42) for Qwen/Qwen3-4B-Instruct-2507, used as a Vietnamese tool-calling agent.
Trained on the LifeMate train split minus a 200-sample dev hold-out, with 50% of user turns replaced by Qwen3-ASR-0.6B transcripts.
| Setup | Input text | Dev score (official metric, 200 samples) |
|---|---|---|
| Qwen3-4B base, prompt + guard + 4-shot | Qwen3-ASR-0.6B | 25.14 |
| Qwen3-4B + this LoRA | PhoWhisper-large | 34.71 |
System budget: Qwen3-4B (4.02B) + PhoWhisper-large (1.55B) = 5.57B parameters.
Serve with vLLM: --enable-lora --lora-modules lifemate=<path> --max-lora-rank 16.