Instructions to use Treese/RQwen3-751M-Base with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Treese/RQwen3-751M-Base with Transformers:

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

pipe = pipeline("text-generation", model="Treese/RQwen3-751M-Base")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Treese/RQwen3-751M-Base")
model = AutoModelForCausalLM.from_pretrained("Treese/RQwen3-751M-Base", 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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Treese/RQwen3-751M-Base with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Treese/RQwen3-751M-Base"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Treese/RQwen3-751M-Base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Treese/RQwen3-751M-Base

SGLang

How to use Treese/RQwen3-751M-Base 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 "Treese/RQwen3-751M-Base" \
    --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": "Treese/RQwen3-751M-Base",
		"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 "Treese/RQwen3-751M-Base" \
        --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": "Treese/RQwen3-751M-Base",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Treese/RQwen3-751M-Base with Docker Model Runner:
```
docker model run hf.co/Treese/RQwen3-751M-Base
```

Untied Qwen3 vocabulary tax and depth allocation on 751M budget

by AndrewThompson1233 - opened about 16 hours ago

Discussion

AndrewThompson1233

about 16 hours ago

Hi Treese,

Pretraining a 751M Qwen3 base model from scratch on 13B tokens across educational web corpora is a solid engineering effort, and logging the benchmark reality on ARC/MMLU is refreshing.

Looking at your architecture parameters:
Vocabulary size is 151,936 at hidden width 1024.
Untying the LM head means holding two full tables: 151,936 * 1024 * 2 = 311.16M parameters.
Lookup tables take 41.4% of your total 751.6M budget, leaving 440.4M for active sequence modeling across 28 layers.
With GQA (16:8) and SwiGLU (intermediate 3072), each transformer block costs roughly 15.73M parameters.

In an open architecture project called Maba (101M reference model: https://huggingface.co/AndrewThompson1233/maba-v1-architecture), we decouple large vocabulary tables from hidden dimensions using low-rank projection:
Projecting 151,936 -> 128 -> 1024 drops table size from 155.58M to ~19.58M parameters.
Even in an asymmetric setup (factorizing only the input table to avoid the output softmax bottleneck while keeping the LM head full rank), you reclaim ~136M parameters.
That saved budget funds 8 to 9 additional Qwen3 layers (expanding depth from 28 to 36-37 layers) within the exact same 751M ceiling and FLOP footprint.

Given that ARC-Challenge and MMLU struggle primarily with multi-step relational depth at this scale, did you benchmark tied vs untied heads before locking in the 28-layer depth?

Best,
Andrew

Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.

Tap or paste here to upload images

· Sign up or log in to comment