Instructions to use grenishrai/suzu-1.7b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use grenishrai/suzu-1.7b with Transformers:

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

pipe = pipeline("text-generation", model="grenishrai/suzu-1.7b")
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("grenishrai/suzu-1.7b")
model = AutoModelForCausalLM.from_pretrained("grenishrai/suzu-1.7b", 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 grenishrai/suzu-1.7b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "grenishrai/suzu-1.7b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "grenishrai/suzu-1.7b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/grenishrai/suzu-1.7b

SGLang

How to use grenishrai/suzu-1.7b 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 "grenishrai/suzu-1.7b" \
    --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": "grenishrai/suzu-1.7b",
		"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 "grenishrai/suzu-1.7b" \
        --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": "grenishrai/suzu-1.7b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Unsloth Desktop
Docker Model Runner
How to use grenishrai/suzu-1.7b with Docker Model Runner:
```
docker model run hf.co/grenishrai/suzu-1.7b
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Suzu-1.7B

Suzu is the second-generation model in the Mori family by Grenish Rai: a QLoRA SFT of HuggingFaceTB/SmolLM2-1.7B for reasoning, math, coding, and technical help, merged to fp16. Its predecessor is Yoru.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tok = AutoTokenizer.from_pretrained("grenishrai/Suzu-1.7B")
model = AutoModelForCausalLM.from_pretrained("grenishrai/Suzu-1.7B", torch_dtype=torch.float16, device_map="auto")
msgs = [{"role": "user", "content": "A train travels 120 km in 1.5 hours. What is its average speed?"}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
out = model.generate(ids, max_new_tokens=256, do_sample=False)
print(tok.decode(out[0, ids.shape[-1]:], skip_special_tokens=True))

The chat template injects a short Suzu system prompt when none is given. Training also used the full Suzu prompt and no prompt at all, so all three work. eos is <|im_end|>, pad is <empty_output>.

Training

Setting	Value
Base	`HuggingFaceTB/SmolLM2-1.7B` (base checkpoint; ChatML embeddings mean-initialized before training)
Method	QLoRA (4-bit NF4) with Unsloth on one Kaggle T4, completion-only loss, merged to fp16
LoRA	r=32, alpha=64, dropout 0, all attention + MLP projections
Data	`grenishrai/math-sft-dataset` 70% train split (61,230 rows incl. identity set)
System prompt mix	40% full / 40% short / 20% none
Steps	5452 (1.42 epochs), global batch 16, max length 2048
LR	0.0002, cosine, 3% warmup, adamw_8bit
Checkpoint kept	best validation loss (0.4768)

Evaluation

Metric	Value
Held-out test loss (1,000-row sample of the 20% split)	0.5259
GSM8K test, 250-problem sample, 0-shot greedy (4-bit + LoRA, before merge)	47.2%
GSM8K test, 64-problem subset, 0-shot greedy (these merged fp16 weights)	48.4%

The GSM8K numbers come from samples with a simple final-number match, not an official harness run.

Limitations

1.7B parameters: knowledge benchmarks (MMLU, ARC, trivia) are bounded by pretraining, and SFT does not move them much.
Math answers can be wrong while looking confident. Verify important results.
The training mix is math-heavy (GSM8K, MetaMathQA) with an OpenHermes-2.5 general slice. Its answers are machine-generated and can contain errors.
No safety tuning beyond the SFT data.

Downloads last month: 23

Safetensors

Model size

2B params

Tensor type

BF16

Model tree for grenishrai/suzu-1.7b

Base model

HuggingFaceTB/SmolLM2-1.7B

Finetuned

(63)

this model

Quantizations

1 model

grenishrai
/

suzu-1.7b