# Usage

Load `horiuchinobuyuki/Qwick-3.5-9B` as a normal Transformers checkpoint. The validated environment used Python 3.11.15, PyTorch 2.11.0+cu130, Transformers 5.14.1, Accelerate 1.12.0, and Safetensors 0.8.0.

```python
import torch
from transformers import AutoTokenizer, Qwen3_5ForConditionalGeneration

model_id = "horiuchinobuyuki/Qwick-3.5-9B"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = Qwen3_5ForConditionalGeneration.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
).eval()

inputs = tokenizer.apply_chat_template(
    [{"role": "user", "content": "Explain why the harmonic series diverges."}],
    tokenize=True,
    add_generation_prompt=True,
    enable_thinking=True,
    return_tensors="pt",
    return_dict=True,
).to(next(model.parameters()).device)

with torch.inference_mode():
    output = model.generate(
        **inputs,
        max_new_tokens=8192,
        do_sample=True,
        temperature=1.0,
        top_p=0.95,
        top_k=20,
        min_p=0.0,
        repetition_penalty=1.0,
    )

completion = output[0, inputs.input_ids.shape[-1]:]
print(tokenizer.decode(completion, skip_special_tokens=True))
```

The example uses an 8,192-token cap. The MMLU-Pro and GPQA runs used vLLM with `presence_penalty=1.5` and a 32,768-token cap; exact profiles are in `evaluation_results.json`. Backend, quantization, context, and decoding affect accuracy and length.

Set `enable_thinking=False` for non-thinking generation. This mode is inherited from Qwen3.5, but the checkpoint was selected primarily with thinking enabled.

The BF16 weight files total 18.82 GB (17.53 GiB). A 24 GB GPU is a practical starting point for short prompts; context capacity depends on the backend and KV cache. Multimodal inputs are supported by the architecture. A matched BF16 measurement over all 900 MMMU validation rows is reported in `EVALUATION.md`, but the validated usage example and Hub pipeline remain text-generation focused.