MMLU-Pro (full test split, 12,032 questions), generative 5-shot CoT, run with the Qwen3-30B-A3B-Instruct-2507 sampling recipe (temperature=0.7, top_p=0.8, top_k=20, min_p=0) via lm-eval-harness on vLLM.

Model	MMLU-Pro
Original	77.88%
Pearl	77.33%

Intended Use

This model is intended to be served through the Pearl miner stack, where vLLM inference is integrated with Pearl mining workflows.

Typical flow:

Run pearld with RPC enabled.
Start the Pearl miner/vLLM stack.
Serve this model through vLLM while Pearl gateway/miner components handle mining-side integration.

How To Use (Pearl vLLM Plugin)

Follow the miner setup from the Pearl repo:

pearl/miner README

High-level prerequisites:

Python 3.12
uv
CUDA + NVIDIA GPU (sm90 class, e.g. H100/H200, per project docs)
Rust toolchain
Running pearld node with RPC credentials

Docker Example

From the Pearl repository root:

docker buildx build -t vllm_miner . -f miner/vllm-miner/Dockerfile

bash

docker run --rm -it --gpus all \
  -p 8000:8000 -p 8337:8337 -p 8339:8339 \
  -e PEARLD_RPC_URL=<PEARLD_URL> \
  -e PEARLD_RPC_USER=<RPC_USER> \
  -e PEARLD_RPC_PASSWORD=<RPC_PASSWORD> \
  -e VLLM_USE_DEEP_GEMM=0 \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  --shm-size 8g \
  vllm_miner:latest \
  pearl-ai/Qwen3-30B-A3B-Instruct-2507-pearl \
  --host 0.0.0.0 --port 8000 \
  --max-model-len 8192 \
  --gpu-memory-utilization 0.9 \
  --enforce-eager

vLLM Standalone

bash

pip install vllm

vllm serve "pearl-ai/Qwen3-30B-A3B-Instruct-2507-pearl"

bash

curl -X POST "http://localhost:8000/v1/chat/completions" \
  -H "Content-Type: application/json" \
  --data '{
    "model": "pearl-ai/Qwen3-30B-A3B-Instruct-2507-pearl",
    "messages": [
      {
        "role": "user",
        "content": "What is the capital of France?"
      }
    ]
  }'

Transformers

python

from transformers import pipeline

pipe = pipeline("text-generation", model="pearl-ai/Qwen3-30B-A3B-Instruct-2507-pearl")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

python

from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("pearl-ai/Qwen3-30B-A3B-Instruct-2507-pearl")
model = AutoModelForCausalLM.from_pretrained("pearl-ai/Qwen3-30B-A3B-Instruct-2507-pearl")

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]:]))

License

This model is based on Qwen3 and is distributed under the Apache 2.0 license.

Limitations

Can generate incorrect, unsafe, or biased outputs.
Requires careful deployment controls and output validation.
Hardware/software compatibility depends on the Pearl miner stack and supported GPU architectures.