Instructions to use endishai/qwen2.5-32b-lexenvs-grpo with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use endishai/qwen2.5-32b-lexenvs-grpo with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="endishai/qwen2.5-32b-lexenvs-grpo")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("endishai/qwen2.5-32b-lexenvs-grpo")
model = AutoModelForCausalLM.from_pretrained("endishai/qwen2.5-32b-lexenvs-grpo")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Inference
Notebooks
Google Colab
Kaggle
Local Apps

vLLM

How to use endishai/qwen2.5-32b-lexenvs-grpo with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "endishai/qwen2.5-32b-lexenvs-grpo"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "endishai/qwen2.5-32b-lexenvs-grpo",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/endishai/qwen2.5-32b-lexenvs-grpo

SGLang

How to use endishai/qwen2.5-32b-lexenvs-grpo with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "endishai/qwen2.5-32b-lexenvs-grpo" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "endishai/qwen2.5-32b-lexenvs-grpo",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "endishai/qwen2.5-32b-lexenvs-grpo" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "endishai/qwen2.5-32b-lexenvs-grpo",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use endishai/qwen2.5-32b-lexenvs-grpo with Docker Model Runner:
```
docker model run hf.co/endishai/qwen2.5-32b-lexenvs-grpo
```

Qwen2.5-32B LexEnvs GRPO

A GRPO-trained variant of Qwen/Qwen2.5-32B-Instruct specialized for credit card optimization reasoning.

On a held-out test set of 30 tasks, this model scores ~0.51 average reward, outperforming Claude Opus 4.6 (~0.41), Claude Sonnet 4.6 (0.396), and GPT-4o (0.363).

Evaluation

Model	Test Avg Reward
Qwen 32B (base)	~0.24
GPT-4o	0.363
Claude Sonnet 4.6	0.396
Claude Opus 4.6	~0.41
This model (Qwen 32B + GRPO)	~0.51

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "endishai/qwen2.5-32b-lexenvs-grpo", torch_dtype="auto", device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained("endishai/qwen2.5-32b-lexenvs-grpo")

messages = [
    {"role": "system", "content": "You are a financial advisor..."},
    {"role": "user", "content": "I spend $600/mo on dining..."},
]
inputs = tokenizer.apply_chat_template(messages, return_tensors="pt", add_generation_prompt=True).to(model.device)
outputs = model.generate(inputs, max_new_tokens=4096, temperature=0.4)
print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))

Details

Base model: Qwen/Qwen2.5-32B-Instruct
Training: GRPO (via TRL) with LoRA rank 32 on 2× A100-80GB
Training data: endishai/lexenvs-tasks
LoRA adapter only: endishai/qwen2.5-32b-lexenvs-grpo-lora
License: Apache 2.0

Intended Use

Credit card optimization reasoning and portfolio selection. Not for live consumer financial advice.

Citation

@misc{lexenvs2026,
  title={LexEnvs: A Harbor RL Environment for Credit Card Optimization},
  author={Imberman, Daniel and Book, Kenny},
  year={2026},
  url={https://huggingface.co/endishai/qwen2.5-32b-lexenvs-grpo}
}

Downloads last month: 4

Safetensors

Model size

33B params

Tensor type

BF16

Model tree for endishai/qwen2.5-32b-lexenvs-grpo

Base model

Qwen/Qwen2.5-32B

Finetuned

Qwen/Qwen2.5-32B-Instruct

Finetuned

(1218)

this model

endishai
/

qwen2.5-32b-lexenvs-grpo