Llama-3.2-Mavrick.Spark-1B

By 11-47 / WithinUsAI

A fast, edge-friendly 1.23B parameter model that distills the Spark-style persona (concise, helpful, agentic vibes) and Maverick long-form reasoning traces into Meta's Llama 3.2 1B Instruct base.

Built with Llama 3.2. This is a community research distillation. Not affiliated with Meta. Llama 3.2 is licensed under the Llama 3.2 Community License.

Model Details

• Parameters: 1.23B
• Context Length: 128k
• Model Size: ~2.49 GB (FP16 / Safetensors)
• Base Model: meta-llama/Llama-3.2-1B-Instruct
• Tensor Type: FP16

Training Data

• Dataset 1: WithinUsAI/Meta_Muse_Spark_Distilled_5k — 5k Spark-style instructions, persona prompts, and tool-use traces.
• Dataset 2: WithinUsAI/Llama_4_Maverick_Distilled_5k — 5k long-form Chain-of-Thought / reasoning traces from Maverick.
• Total: 10k high-quality distilled conversations.
• Method: Supervised Fine-Tuning (SFT) on assistant tokens only (ShareGPT / ChatML format with system + user + assistant roles).

Why This Model Rocks

Llama 3.2 1B is perfect for phones, laptops, and edge devices but can feel shallow on reasoning. This distillation compresses the clever, agentic flavor of Muse Spark (Meta's closed frontier model) and Llama 4 Maverick into a runnable local model. Great for quick inference while keeping a spark of that frontier reasoning style.

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "11-47/Llama-3.2-Mavrick.Spark-1B"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    device_map="auto",
    torch_dtype="auto"
)

messages = [
    {"role": "system", "content": "You are Mavrick.Spark, a concise, helpful assistant built by WithinUsAI."},
    {"role": "user", "content": "Explain how distillation works in 3 bullets."}
]

inputs = tokenizer.apply_chat_template(messages, add_generation_prompt=True, return_tensors="pt").to(model.device)

outputs = model.generate(
    inputs,
    max_new_tokens=400,
    temperature=0.7,
    top_p=0.9,
    do_sample=True
)

print(tokenizer.decode(outputs[0][inputs.shape[1]:], skip_special_tokens=True))
The repo includes a chat_template.jinja, so chat interfaces and HF Inference work out of the box.
Intended Uses
    •	Local / on-device inference (phones, laptops, edge devices)
    •	Conversational agents with Spark-like personality
    •	Research on knowledge distillation & style transfer
    •	Lightweight reasoning tasks
Limitations
    •	Still a 1B model — will hallucinate more than full-size Maverick/Spark models.
    •	Distilled style can sometimes over-explain.
    •	Knowledge cutoff inherited from base Llama 3.2 (~Dec 2023) + synthetic traces.
    •	Not production-hardened for heavy tool use or code execution.
License & Attribution
    •	Base weights: Llama 3.2 Community License — keep the “Llama” prefix and display “Built with Llama”.
    •	Datasets are from WithinUsAI.
    •	Please cite the original repo and base model if you build on this.
Citation
@misc{llama-3.2-mavrick-spark-1b,
  author = {11-47 / WithinUsAI},
  title = {Llama-3.2-Mavrick.Spark-1B},
  year = {2026},
  publisher = {Hugging Face},
  url = {https://huggingface.co/11-47/Llama-3.2-Mavrick.Spark-1B}
}