Instructions to use ifx-pse-sys-ml/spark-13m-instruct with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ifx-pse-sys-ml/spark-13m-instruct with Transformers:

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

pipe = pipeline("text-generation", model="ifx-pse-sys-ml/spark-13m-instruct", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("ifx-pse-sys-ml/spark-13m-instruct", trust_remote_code=True, device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use ifx-pse-sys-ml/spark-13m-instruct with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ifx-pse-sys-ml/spark-13m-instruct"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ifx-pse-sys-ml/spark-13m-instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ifx-pse-sys-ml/spark-13m-instruct

SGLang

How to use ifx-pse-sys-ml/spark-13m-instruct 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 "ifx-pse-sys-ml/spark-13m-instruct" \
    --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": "ifx-pse-sys-ml/spark-13m-instruct",
		"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 "ifx-pse-sys-ml/spark-13m-instruct" \
        --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": "ifx-pse-sys-ml/spark-13m-instruct",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use ifx-pse-sys-ml/spark-13m-instruct with Docker Model Runner:
```
docker model run hf.co/ifx-pse-sys-ml/spark-13m-instruct
```

spark-13m-instruct

A 13.2M-parameter English instruction-tuned language model — a deliberately tiny model in the spirit of SmolLM, for research on small-model capability, fast experimentation, and as a lightweight decoder backbone (e.g. for a small VLM). It is ~10× smaller than SmolLM-135M.

Architecture: Llama-style decoder, hidden 384, 6 layers, 6 heads / 2 KV heads (GQA), intermediate 1216, RoPE, context 1024, vocab 6400 (English BPE).
Base: pretrained on ~17B tokens (TinyStories + FineWeb-Edu + ClimbMix), wiki perplexity 21.98.
Instruct: SFT on SmolTalk (200k conversations), 5 epochs.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("ifx-pse-sys-ml/spark-13m-instruct", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained("ifx-pse-sys-ml/spark-13m-instruct")

msgs = [{"role": "user", "content": "Tell me about the moon."}]
ids = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt")
out = model.generate(ids, max_new_tokens=128)
print(tok.decode(out[0][ids.shape[1]:], skip_special_tokens=True))

The decoder also accepts inputs_embeds (exactly one of input_ids / inputs_embeds), so a vision projector can splice visual tokens in — it can serve as a small VLM text backbone.

Benchmarks

Accuracy (%) via lm-evaluation-harness 0.4, the same harness and shots for every model, so the columns are directly comparable. (SmolLM's blog used lighteval; those numbers use a different harness and are not comparable, so they are not quoted here.)

Benchmark	chance	spark-13m-base	spark-13m-instruct	SmolLM-135M	SmolLM-135M-Instruct
hellaswag	25	28.0	28.3	42.6	41.9
arc_easy	25	38.3	34.2	56.1	43.9
arc_challenge	25	21.4	24.7	28.9	27.4
piqa	50	60.4	56.7	68.4	67.0
winogrande	50	50.1	50.6	53.2	51.3
openbookqa	25	25.4	26.0	34.0	33.6
commonsense_qa	20	21.5	22.4	19.8	20.3
mmlu	25	24.2	25.3	25.2	24.4
average	—	33.7	33.5	41.0	38.7

Honest limitations

At 13M parameters this model is near random chance on knowledge- and harder-reasoning benchmarks (MMLU, OpenBookQA, ARC-Challenge, Winogrande). It retains real signal only where commonsense/pattern beats stored knowledge (PIQA 60, ARC-Easy 38). It is a research and prototyping model — a capable small decoder for its size, not a knowledge model. Instruction-tuning adds response format, not facts, so instruct ≈ base on these academic benchmarks.

Trained with the Nexus codebase. English only.

Downloads last month: -

Safetensors

Model size

13.2M params

Tensor type

F32