Model reference · open weights

SparkAI-llama

Available as managed deployment LLMs vedantjadhav701 · community Text gen 1 variants 653 dl/mo

SparkAI-llama is an open-weight language model from vedantjadhav701. AxForge deploys and operates it for you on dedicated EU-owned hardware — with the licence handled where one is required.

Available as managed deployment — configured and operated for you on dedicated EU hardware, quoted per deployment.

What it is

Released byvedantjadhav701
TypeLanguage models
TaskText gen
Parameters (lead)48M
Context1k tokens
Runs withtransformers
Based onvedantjadhav701/SparkAI-47m-llama-10b-token
Released2026-08-27
Popularity653 downloads / month
LicenceOpen weights

About

What SparkAI-llama is

Instruction-tuned checkpoint of SparkAI-47M-Llama (~48M parameter decoder-only transformer), fine-tuned for chat and instruction following.

🌟 Key Highlights & Unique Features

  • Ultra-Low Memory Footprint (~95.4 MB): Fits in under 100MB of RAM, making it suitable for edge devices, mobile apps, WebGPU, and microcontrollers.
  • 🏋️ Data-Saturated Pretraining (10 Billion Tokens): Pretrained on 10B tokens (210 tokens/param) of high-quality FineWeb-Edu + Cosmopedia-v2 text, providing an empirical benchmark on capacity saturation for sub-50M models.
  • 🏗️ Modern LLaMA 3 Architecture: Built with Grouped Query Attention (GQA), SwiGLU activations, RoPE positional encodings, RMSNorm pre-normalization, and tied embeddings.
  • 💬 Full ChatML SFT Alignment: Fine-tuned with ChatML `` instruction formatting and template support (chat_template.jinja).

Read the full model card

📐 Architecture

  • Parameters: ~48M (~47.4M non-embedding / tied embeddings)
  • Layers: 8
  • Hidden Size: 512
  • Attention Heads: 8 query heads, 2 key/value heads (Grouped Query Attention - GQA)
  • MLP: SwiGLU (Intermediate size: 1408)
  • Positional Encoding: RoPE (Rotary Position Embeddings)
  • Normalization: RMSNorm (Pre-normalization)
  • Embeddings: Tied embeddings, no bias terms
  • Vocabulary Size: 49,152 (SmolLM2 tokenizer with Chat Template)
  • Sequence Length: 1024

🏋️ Training & Fine-Tuning Details

  • Base Checkpoint: vedantjadhav701/SparkAI-47m-llama-10b-token
  • Pretraining Data: FineWeb-Edu (sample-100BT) + Cosmopedia-v2 (85% / 15% mix, 10.00B tokens)
  • Optimizer: AdamW with Cosine LR schedule + warmup
  • Hardware: NVIDIA A100 80GB PCIe
  • Final Eval Perplexity: 31.46

📈 Pretraining Progression

TokensPerplexity
630M43.49
3.77B31.30
7.00B
10.00B31.46

Note on Saturation: Perplexity plateaued between 3.77B and 10.00B tokens despite continued training, indicating the model has saturated its representational capacity at this size.


📊 Comparison in Sub-50M Parameter Landscape

FeatureTypical Sub-50M ModelsSparkAI-47M-Llama / Instruct
Token Budget~1B – 2B tokens10.00 Billion Tokens (210 tokens/param)
Data QualityRaw web text / C4FineWeb-Edu (85%) + Cosmopedia-v2 (15%)
ArchitectureBasic MHA / GPT-2 styleModern LLaMA 3 (GQA, SwiGLU, RoPE, RMSNorm)
Model Size~100MB – 200MB~95.4 MB (model.safetensors)
SFT AlignmentRare / NoneInstruction-tuned with ChatML (chat_template.jinja)
BenchmarkingFew metricsEmpirical capacity saturation documented at 10B tokens

💻 Usage with Hugging Face transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "vedantjadhav701/SparkAI-47m-llama-instruct"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(repo_id)

messages = [
    {"role": "user", "content": "What is a computer program?"}
]

prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=80, do_sample=True, temperature=0.6)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

🚀 Quickstart & Local Gradio UI

Installation

pip install -r requirements.txt

Run Gradio App

python app.py

Open http://127.0.0.1:7860 in your web browser.

From the published model card. Full card on the HuggingFace links in the sidebar.

Using it via the API

Call it like any OpenAI endpoint

Once AxForge deploys sparkai-llama for you, it answers on the OpenAI-compatible API — the same base URL and keys as every other model. (sparkai-llama below is illustrative; you get the exact model name on deployment.)

$ curl -sS https://api.axforge.ai/v1/chat/completions \
  -H "Authorization: Bearer $AXFORGE_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"model":"sparkai-llama","messages":[{"role":"user","content":"Hello"}]}'

Create an account — your API key is available in the console. 3M free tokens every 30 days with every new account.

© 2026 AxForge · EU-hosted AI infrastructure Pricing Docs Trust Privacy Terms