SPADE-Qwen3-4B-Games

SPADE-Qwen3-4B-Games

SPADE checkpoint for the games setting, trained from Qwen/Qwen3-4B-Instruct-2507.

SPADE trains a single model in two roles: an Environment Designer that writes executable environments, and a Reasoning Agent that solves them. The Designer is rewarded for producing environments at the frontier of what the Agent can currently solve, so the curriculum keeps pace with the policy instead of being fixed in advance. See the paper for details.

Quickstart

We advise you to use the latest version of transformers.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "spade-rl/SPADE-Qwen3-4B-Games"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

prompt = "Give me a short introduction to large language model."
messages = [{"role": "user", "content": prompt}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
model_inputs = tokenizer([text], return_tensors="pt").to(model.device)

generated_ids = model.generate(**model_inputs, max_new_tokens=16384)
output_ids = generated_ids[0][len(model_inputs.input_ids[0]):].tolist()
print(tokenizer.decode(output_ids, skip_special_tokens=True))

Deployment

For deployment, you can use sglang>=0.4.6.post1 or vllm>=0.8.5 to create an OpenAI-compatible API endpoint:

• SGLang:

  python -m sglang.launch_server --model-path spade-rl/SPADE-Qwen3-4B-Games --context-length 262144
  

• vLLM:

  vllm serve spade-rl/SPADE-Qwen3-4B-Games --max-model-len 262144
  

Note: If you encounter out-of-memory (OOM) issues, consider reducing the context length to a shorter value, such as 32,768.

Best practices

We recommend temperature=0.7, top_p=0.8, top_k=20, min_p=0, following the sampling guidance on the base model card.

Related artifacts

• Grounding corpora: games | tool use
• All SPADE models, data and generated environments: huggingface.co/spade-rl