Qwen3.5-27B-Claude-Opus-4.6-High-Reasoning

A distilled version of Qwen3.5-27B, enhanced with Claude Opus 4.6 reasoning patterns through knowledge distillation.

Evaluation Results

• Evaluator: Qwen3-Coder-Next
• Test Samples: 260 (based on personal dataset, for reference only)
• Date: 2026-03-29

Overall Performance

Metric	Base Model	Distilled Model	Improvement
Win Rate	25.77%	73.85%	+48.08%
Avg Latency (s)	71.70	68.47	-4.5%

Score Breakdown (10-point scale)

Dimension	Base Model	Distilled Model	Improvement
Accuracy	6.35	8.59	+35.3%
Logic	6.47	8.69	+34.3%
Completeness	5.70	8.82	+54.7%
Clarity	6.51	8.43	+29.5%
Actionability	5.85	8.56	+46.3%

Performance by Category

Category	Base Win Rate	Distilled Win Rate	Samples
debug	22.67%	76.00%	75
design	51.32%	48.68%	76
prompt	5.80%	94.20%	69
reasoning	17.50%	82.50%	40

Quick Start with vLLM

Installation

pip install vllm

Online Inference (Simple)

from vllm import LLM, SamplingParams

# Initialize the model
llm = LLM(
    model="HarleyWang/Qwen3.5-27B-Claude-Opus-4.6-High-Reasoning",
    dtype="bfloat16",
    tensor_parallel_size=2,  # Adjust based on your GPU count
    max_model_len=8192,
)

# Define sampling parameters
sampling_params = SamplingParams(
    temperature=0.7,
    top_p=0.9,
    max_tokens=2048,
)

# Create messages
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Explain how to optimize a slow database query."}
]

# Use apply_chat_template from tokenizer
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("HarleyWang/Qwen3.5-27B-Claude-Opus-4.6-High-Reasoning")
prompt = tokenizer.apply_chat_template(messages, tokenize=False)

# Generate
outputs = llm.generate(prompt, sampling_params)

# Print output
for output in outputs:
    print(output.outputs[0].text)

Online Inference (Streaming)

from vllm import LLM, SamplingParams
from transformers import AutoTokenizer

llm = LLM(
    model="HarleyWang/Qwen3.5-27B-Claude-Opus-4.6-High-Reasoning",
    dtype="bfloat16",
    tensor_parallel_size=2,
)

tokenizer = AutoTokenizer.from_pretrained("HarleyWang/Qwen3.5-27B-Claude-Opus-4.6-High-Reasoning")

messages = [
    {"role": "user", "content": "Write a Python function to sort a list."}
]
prompt = tokenizer.apply_chat_template(messages, tokenize=False)

sampling_params = SamplingParams(
    temperature=0.7,
    top_p=0.9,
    max_tokens=1024,
    stream=True,  # Enable streaming
)

for output in llm.generate(prompt, sampling_params, use_tqdm=False):
    print(output.outputs[0].text, end="", flush=True)
print()

Offline Inference (Batch)

from vllm import LLM, SamplingParams

llm = LLM(
    model="HarleyWang/Qwen3.5-27B-Claude-Opus-4.6-High-Reasoning",
    dtype="bfloat16",
    tensor_parallel_size=2,
)

prompts = [
    "Explain quantum computing.",
    "Write a haiku about coding.",
    "Debug: Why is this loop infinite?",
]

sampling_params = SamplingParams(
    temperature=0.7,
    top_p=0.9,
    max_tokens=512,
)

outputs = llm.generate(prompts, sampling_params)

for output in outputs:
    print(f"Prompt: {output.prompt}")
    print(f"Response: {output.outputs[0].text}")
    print("-" * 50)

Command Line Interface

# Start an OpenAI-compatible API server
python -m vllm.entrypoints.openai.api_server \
    --model HarleyWang/Qwen3.5-27B-Claude-Opus-4.6-High-Reasoning \
    --dtype bfloat16 \
    --tensor-parallel-size 2 \
    --host 0.0.0.0 \
    --port 8000

# In another terminal, use the API
curl http://localhost:8000/v1/chat/completions \
    -H "Content-Type: application/json" \
    -d '{
        "model": "default",
        "messages": [
            {"role": "system", "content": "You are a helpful assistant."},
            {"role": "user", "content": "What is the capital of France?"}
        ],
        "temperature": 0.7,
        "max_tokens": 256
    }'

License

This model is licensed under Apache-2.0.

---

This model was distilled to capture Claude Opus 4.6's reasoning capabilities while maintaining Qwen3.5-27B's efficiency and multilingual support. Recommended for inference with vLLM for optimal performance.