GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
dokster
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
dokster
Model Tree
Qwen/Qwen3.5-2B
Input Modalities
Output Modalities
Supported Functionality
| rank / alpha / dropout | 16 / 32 / 0.1 |
| lr / schedule | 0.0001 cosine, warmup 0.05 |
| batch / epochs | 8 x 1, up to 3 (early stopping on eval loss) |
| max seq len | 128 |
| targets | every Linear in the language model (attention + MLP + Gated DeltaNet projections) |
python
import torchfrom transformers import AutoTokenizer, AutoModelForCausalLMfrom peft import PeftModeltok = AutoTokenizer.from_pretrained("dokster/qwen3.5-jason-statham-lora")base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3.5-2B", dtype=torch.bfloat16, device_map="auto")model = PeftModel.from_pretrained(base, "dokster/qwen3.5-jason-statham-lora")prompt = tok.apply_chat_template([{"role": "system", "content": "Ты — Джейсон Стэйтем. Отвечай короткой жёсткой цитатой."},{"role": "user", "content": "Цитата:"}],tokenize=False, add_generation_prompt=True, enable_thinking=False,)enc = tok(prompt, return_tensors="pt", add_special_tokens=False).to(model.device)out = model.generate(**enc, max_new_tokens=64, do_sample=True, temperature=0.95, top_p=0.95)print(tok.decode(out[0][enc["input_ids"].shape[1]:], skip_special_tokens=True))
enable_thinking=False is not optional: the adapter was trained on the closed,
empty <think> block, and the default generation prompt leaves it open.