Model reference · open weights

Supra

LLMs SupraLabs Text gen 1 build Open weights 1k dl/mo

Supra is an open-weight language model from SupraLabs. Supra-50M-Base (FP32) weighs 104 MB; the smallest configuration that runs it is RTX 3060 12 GB.

For self-hosting, compare this build's requirements below with the GPU you plan to rent.

  • Supra-50M is a 52M-parameter causal language model developed by SupraLabs for text generation tasks.
  • It utilizes a Llama-style architecture with a context length of 1024 tokens and supports English.
  • The model is trained on 20 billion tokens of educational web text and is released under the Apache 2.0 license.

Summary of the SupraLabs/Supra-50M-Base model card, 2026-10-01

What it is

Released bySupraLabs
Released2026-05-21
Parameters52M
VRAM104 MB for the weights

What it runs on

Memory and cards for Supra-50M-Base (FP32)

104 MBweights, file size
12 MBcache per 1K tokens
482 MBruntime overhead, at least
1,024 tokenscontext max
CardRequests at onceContext maxMemory
1K each
RTX 3060 12 GB876all 1K11.6 GB
RTX 4060 Ti 16 GB1000+all 1K15.4 GB
RTX 3090 24 GB1000+all 1K23.4 GB
RTX 4090 24 GB1000+all 1K23.4 GB
RTX 5090 32 GB1000+all 1K31.0 GB
L40S 48 GB1000+all 1K44.0 GB
A100 80 GB1000+all 1K78.2 GB
H100 80 GB1000+all 1K78.1 GB
RTX PRO 6000 Blackwell 96 GB1000+all 1K93.8 GB
DGX Spark (GB10) 128 GB unified1000+all 1K107 GB
H200 141 GB1000+all 1K138 GB
B200 180 GB1000+all 1K176 GB
Memory needed at each load
Requests at once1K tokens each
1598 MB
5649 MB
8686 MB
16787 MB
32988 MB
641.4 GB

One card, with vLLM's small-card settings.

From the model card

What SupraLabs says about Supra

Read the model card

Supra-50M is a compact 50M-parameter BASE causal language model built by SupraLabs, trained from scratch using a Llama-style architecture on 20 billion tokens of high-quality educational web text. Despite being significantly smaller than comparable open models, it achieves competitive or superior results on several key benchmarks. It's our first SupraLabs Scaling Up Plan model.


🏆 Benchmarks

BenchmarkSupra-50M (ours)GPT-2 (124M)SmolLM-135MOpenELM-270M
Parameters50M124M (2.5×)135M (2.7×)270M (5.4×)
BLiMP (linguistics)76.3%63.0%69.8%(N/A)
SciQ (science)77.2%53.2%73.4%84.70%
ARC-Easy (knowledge)52.2%42.0%49.2%45.08%
PIQA (logic)62.2%63.0%67.3%69.75%
HellaSwag (context)31.8%29.5%42.0%46.71%
TaskMetricValue
arc_easyacc,none0.5185
arc_easyacc_stderr,none0.0103
arc_easyacc_norm,none0.4600
arc_easyacc_norm_stderr,none0.0102
arc_challengeacc,none0.2159
arc_challengeacc_stderr,none0.0120
arc_challengeacc_norm,none0.2517
arc_challengeacc_norm_stderr,none0.0127
hellaswagacc,none0.2903
hellaswagacc_stderr,none0.0045
hellaswagacc_norm,none0.3172
hellaswagacc_norm_stderr,none0.0046
winograndeacc,none0.5154
winograndeacc_stderr,none0.0140
piqaacc,none0.6251
piqaacc_stderr,none0.0113
piqaacc_norm,none0.6219
piqaacc_norm_stderr,none0.0113
openbookqaacc,none0.1860
openbookqaacc_stderr,none0.0174
openbookqaacc_norm,none0.3080
openbookqaacc_norm_stderr,none0.0207
boolqacc,none0.5303
boolqacc_stderr,none0.0087

🧠 Model Architecture & Hyperparameters

Supra-50M is based on the LlamaForCausalLM architecture with the following configuration:

HyperparameterValue
ArchitectureLlama (decoder-only transformer)
Parameters~50M
vocab_size32,000
hidden_size512
intermediate_size1,408
num_hidden_layers12
num_attention_heads8
num_key_value_heads4 (GQA)
max_position_embeddings1,024
rope_theta10,000
tie_word_embeddingsTrue

📚 Training Data

PropertyValue
DatasetHuggingFaceFW/fineweb-edu (sample-100BT split)
Total tokens20,000,000,000 (20B)
Sequence length1,024 tokens
Storage formatMemory-mapped binary (uint16, ~40 GB)

🔤 Tokenizer

A custom Byte-Level BPE tokenizer was trained from scratch on 500,000 documents sampled from fineweb-edu (sample-10BT).

PropertyValue
TypeByteLevelBPETokenizer
Vocabulary size32,000
Min frequency2
Special tokens, , , , ``

⚙️ Training Configuration

ParameterValue
Epochs1
Per-device batch size32
Gradient accumulation steps4
Effective batch size128 × 1,024 tokens
Learning rate6e-4
LR schedulerCosine
Warmup ratio2%
OptimizerAdamW Fused (adam_beta1=0.9, adam_beta2=0.95)
Weight decay0.1
Max grad norm1.0
Precisionbfloat16
torch.compileEnabled
HardwareSingle GPU
Final loss3.259

🚀 Inference

from transformers import pipeline
import torch

print("[*] Loading Supra-50M model from Hugging Face Hub...")
pipe = pipeline(
    "text-generation",
    model="SupraLabs/Supra-50M_BASE",
    device_map="auto",
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
)

def generate_text(prompt, max_new_tokens=150):
    result = pipe(
        prompt,
        max_new_tokens=max_new_tokens,
        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']

# Example
prompt = "The importance of education is"
print(f"\nPrompt: {prompt}")
print("-" * 40)
print("\nOutput:\n" + generate_text(prompt))

💬 Sample Outputs

Prompt: "The main concept of physics is "

The main concept of physics is iffy, and the idea that we can make things behave in a certain way. The most important part of physics is called quantum mechanics which states that all particles are made up of energy (energy) and matter (matter). In physics, there are two types of particles: elementary particles and exotic ones. These particles have properties like mass, speed or momentum but they don’t interact with each other to form new objects. This is because these particles do not exist independently from one another. In this case, an exotic particle might be created by adding more energy into its structure than it would take for a normal particle. However, when you add additional energy to an exotic particle, the new object will become smaller and larger until it becomes too large to fit within the existing structure. If you think about how light travels through space, it takes around 20 billion years before the light reaches our eyes. Light waves travel faster than light at high speeds so if we could create some kind of light wave, then we wouldn’t need any special equipment. It just needs a few hundred millionths of a second to produce light rays. So even though the light is moving along the same path as the current, the speed of light is different depending on where the light hits the

Prompt: "Artificial intelligence is "

Artificial intelligence is iffy, it can be used to make intelligent machines that could take over the world. What does Artificial Intelligence mean? AI refers to ar

Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.

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