We are back. Supra Mini v5 8M is our fifth release in the Supra Mini series and the biggest one yet. Trained on 5 billion tokens of Fineweb-Edu for 2 epochs, v5 jumps to 7.8M parameters and brings real, measurable improvements across every benchmark we ran.

What changed from v4?

v4 had 2.6M parameters and was trained on 3B tokens. v5 triples the parameter count to 7.8M and adds 2 billion more tokens to the training set. The vocab size also doubled from 8,192 to 16,384, giving the model a much richer token vocabulary to work with. Every single part of the config got an upgrade.

// supra mini v5 8m — model config Parameters       â†’ 7,867,584 (~8M)
Architecture    â†’ Llama
Vocab size      â†’ 16,384 (custom BPE)
Hidden size     â†’ 191
Intermediate   â†’ 768
Layers          â†’ 8
Attention heads  â†’ 4
Context length   â†’ 1,024 tokens
Learning rate   â†’ 2e-4
Weight decay    â†’ 0.01
Trained in      â†’ bfloat16

Training setup

We trained v5 on a single NVIDIA RTX 5060 Ti 16GB for approximately 11 hours across 2 epochs. The dataset is the first 5 billion tokens of Sample-10BT from Fineweb-Edu, streamed and tokenized on the fly with our custom BPE tokenizer (vocab size 16,384).

The final training loss after 2 epochs came in at 4.414, down from 4.618 in v4. The full training code is in the repo: train_tokenizer.py, train.py and inference.py.

Benchmarks

We ran the same lm-eval suite as v4 so the numbers are directly comparable. v5 beats v4 on every single task.

Task v5 score v4 score Random baseline
ARC_Easy0.34390.31520.25 (25%)
Wikitext (PPL)2.66173.1652lower is better
BLiMP0.63490.6070.50 (50%)

The Wikitext perplexity drop from 3.16 to 2.66 is the one that stands out most. Lower perplexity means the model is genuinely more confident and accurate at predicting text, not just getting lucky on multiple choice. BLiMP going from 0.607 to 0.634 also shows real grammatical improvement for a model this size.

Example outputs

Same prompts as v4, so you can compare directly. Generation settings: temperature=0.5, top_k=25, top_p=0.9.

// prompt
"Artificial intelligence is "
// output
"Artificial intelligence is _______. I was a scientist at the University of Cambridge in London, and I had been looking for ways to get them back into the first time. It's not just about learning how to use it but also about the world around us. The study has led to more than 150 studies showing that people who are able to learn new things like computers, laptops, and smartphones can be found in many different types of computer systems..."
// prompt
"The main concept of physics is "
// output
"The main concept of physics is _______. This principle, which means that the universe has no energy; it should be a matter of time in its own right and therefore to be able to achieve its maximum potential. It must have been used as an alternative for quantum mechanics but not only in terms of power but also in the form of electrical systems such as magnetic fields or electric currents..."
// prompt
"Once upon a time, "
// output
"Once upon a time, .. ... The word 'pot' is not used in the past. It's like a thing of course that we're going to think about it. The phrase 'these' means what you see and how much are there? You have no idea where I'm so, but if my son has any thought or understanding of his name, he will be able to understand him by saying something more than one day..."

The outputs are noticeably more coherent than v4. Sentences are longer, the topic stays more consistent, and the hallucinations are less random. Still a base model doing base model things, but the progress is clear.

How to run it

from transformers import pipeline
import torch

print("[*] Loading Supra Mini v5 8M model from Hugging Face Hub...")
pipe = pipeline(
    "text-generation",
    model="SupraLabs/Supra-Mini-v5-8M",
    device_map="auto",
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
)

def generate_text(prompt, max_length=150):
    result = pipe(
        prompt,
        max_new_tokens=max_length,
        do_sample=True,
        temperature=0.5,
        top_k=25,
        top_p=0.9,
        repetition_penalty=1.2,
        pad_token_id=pipe.tokenizer.pad_token_id,
        eos_token_id=pipe.tokenizer.eos_token_id
    )
    return result[0]['generated_text']

test_prompt = "The importance of education is"
print(f"\nPrompt: {test_prompt}")
print("-" * 30)
print("\nOutput:\n" + generate_text(test_prompt))

What's next?

v5 is still a base model. No instruction tuning, no chat format, just raw language modeling. The roadmap for the next versions includes fine-tuning experiments and continuing to scale up parameters while keeping everything trainable on consumer hardware.

The model is live. Go try it.

// links Model   â†’ huggingface.co/SupraLabs/Supra-Mini-v5-8M
License → Apache 2.0
Series  â†’ Supra Mini collection on HuggingFace
#release #supra-mini-v5 #tinyml #llama #open-source #fineweb-edu #edge-ai