from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "MananSuri27/gemma4-31b-full-mini-swe-secbench"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(repo_id, dtype="auto", device_map="auto")

vLLM

vllm serve MananSuri27/gemma4-31b-full-mini-swe-secbench --dtype bfloat16

The model was trained on assistant action/tool-call turns from agent trajectories.