Instructions to use hopski/graph-qwenputer with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use hopski/graph-qwenputer with Transformers:

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

pipe = pipeline("text-generation", model="hopski/graph-qwenputer")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("hopski/graph-qwenputer")
model = AutoModelForCausalLM.from_pretrained("hopski/graph-qwenputer", device_map="auto")
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=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use hopski/graph-qwenputer with vLLM:

Install from pip and serve model

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

Use Docker

docker model run hf.co/hopski/graph-qwenputer

SGLang

How to use hopski/graph-qwenputer 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 "hopski/graph-qwenputer" \
    --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": "hopski/graph-qwenputer",
		"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 "hopski/graph-qwenputer" \
        --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": "hopski/graph-qwenputer",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use hopski/graph-qwenputer with Docker Model Runner:
```
docker model run hf.co/hopski/graph-qwenputer
```

Configuration Parsing Warning:Config file tokenizer_config.json cannot be fetched (too big)

Graph Qwenputer: Qwen3.5-4B with a graph computer inside

Qwen3.5-4B with 11 layers added in front of its 32. The added layers are an RV32IM CPU with Zba, compiled by Qwenputer, with 85 added exit units so the machine can print ordinary tokens. Nothing is trained, and Qwen's 424 tensors are bit-identical.

With the project's firmware loaded, the model can read and write an exact graph from inside its own run: typed, directed edges, up to 1,024 nodes, with calls such as path, neighbours, add_node and link. New York's subway is the worked example; the graph can be anything.

These weights hold the computer, and nothing about any graph. Firmware and graph are loaded into the model's running state after the prompt, so the weights alone are an empty machine. They need the driver and firmware in the project repo.

The weights don't change between firmware builds. All 18.76 GB of them are the checkpoint used for the project's real-weights runs.

Limits

The driver offers the tool, loads the memory and routes tokens. Without its routing, a bare top-token loop printed the right answer on 6 of 6 trips I tried, about five times slower, and Qwen's reply afterwards was garbage.
It's slow: a lookup takes minutes where an ordinary tool takes milliseconds, and an ordinary tool gets the same answers.
Qwen's judgment is the weak part. See the repo's known problems.

Credits and license

The compiler and machine are Divyansh Agrawal's Qwenputer (MIT). The base weights are Qwen3.5-4B (Apache 2.0, per its model card); its license ships here as LICENSE.

Downloads last month: 15

Safetensors

Model size

9B params

Tensor type

BF16

F32

Model tree for hopski/graph-qwenputer

Base model

Qwen/Qwen3.5-4B-Base

Finetuned

Qwen/Qwen3.5-4B

Finetuned

(1,292)

this model