Apriel-5B-base is a decoder-only transformer trained on 4.5T+ tokens of data. It is the first release in the Apriel model family, designed to support research on foundation models. Apriel-5B-base achieves strong performance across common benchmarks for models under 5B parameters.

Apriel-5B-Instruct

Apriel-5B-Instruct is built on top of Apriel-5B-base using continual pretraining (CPT), supervised finetuning (SFT), and post-training alignment with DPO and RLVR.

Both CPT and SFT stages involved training multiple domain-biased variants with overlapping datasets (e.g., instruction, code, math). These were then merged to form a more general-purpose model before alignment. The final model is aligned for instruction following, reasoning, and safety-aware dialogue.

The y-axis shows average downstream benchmark scores. Throughput (x-axis) was measured using vLLM with batch size 8, 256 input tokens, and 32 output tokens.

How to Use

pip install transformers

Running the Base model

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

checkpoint = "ServiceNow-AI/Apriel-5B-Base"
device = "cuda"  # or "cpu"

tokenizer = AutoTokenizer.from_pretrained(checkpoint)
model = AutoModelForCausalLM.from_pretrained(checkpoint, torch_dtype=torch.bfloat16).to(device)

inputs = tokenizer.encode("Snow is", return_tensors="pt").to(device)
outputs = model.generate(inputs)
print(tokenizer.decode(outputs[0]))

bash

>>> print(f"Memory footprint: {model.get_memory_footprint() / 1e6:.2f} MB")
Memory footprint: 9664.14 MB

Running the Instruct model

python

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

checkpoint = "ServiceNow-AI/Apriel-5B-Instruct"
tokenizer = AutoTokenizer.from_pretrained(checkpoint)
device = "cuda" if torch.cuda.is_available() else "cpu"

model = AutoModelForCausalLM.from_pretrained(
    checkpoint, 
    torch_dtype=torch.bfloat16 if device == "cuda" else torch.float32
).to(device)

messages = [
    {"role": "system", "content": "You are a helpful AI assistant that provides accurate and concise information."},
    {"role": "user", "content": "Tell me about artificial intelligence"}
]

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

generation_params = {
    "max_new_tokens": 512,
    "temperature": 0.2,
    "top_p": 0.9,
    "do_sample": True
}

outputs = model.generate(**inputs, **generation_params)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Chat Template

<|system|>
System message here (optional)
<|end|>
<|user|>
User message here
<|end|>
<|assistant|>
Assistant response here
<|end|>

If no system message is provided, the model inserts a blank system prompt to maintain format structure. The model supports structured interaction patterns, including tool calling and reasoning steps for more advanced workflows.

Evaluation

Evaluations were conducted using lm-eval-harness and evalchemy.

Apriel-5B-Base

Task Name	Apriel-5B-Base	OLMo-2-1124-7B	Llama-3.1-8B	Mistral-Nemo-Base-2407
Average	58.7	58.71	61.72	66.01
ARC Challenge	56.7	62.7	58.2	62.9
ARC Easy	82.4	86.0	85.7	86.7
MMMLU	44.5	35.3	47.4	54.7
Global MMLU	57.4	52.4	61.1	68.4
GSM8k	64.2	63.2	54.8	58.5
HellaSwag	74.4	80.5	78.8	82.7
MUSR	39.1	39.6	38.0	39.9
MBPP	27.6	22.4	46.0	54.6
MMLU	61.3	63.9	66.0	69.6
PIQA	78.9	81.1	81.2	82.1

Apriel-5B-Instruct

Task Name	Apriel-5B-Instruct	OLMo-2-1124-7B-Instruct	Llama-3.1-8B-Instruct	Mistral-Nemo-Instruct-2407
Average	49.64	43.91	52.60	48.63
ARC Challenge	59.04	61.45	64.25	66.38
GSM8k	80.36	79.68	82.63	77.63
Hellaswag	74.52	80.21	78.43	81.71
BBH	39.82	39.95	50.86	50.06
GPQA	28.36	27.85	29.19	29.45
IF Eval	80.78	72.64	79.67	62.85
MMLU Pro	29.19	26.57	37.74	35.09
MUSR	36.77	34.39	38.36	39.02
MBPP	45.80	28.00	59.00	57.60
TruthfulQA	56.09	56.46	55.05	57.69
Winogrande	62.35	65.35	67.01	70.01
Minerva Math	39.80	9.96	36.72	21.46
MATH500	53.00	31.4	45.80	34.40
AMC23	29.00	16.4	21.00	11.50
MixEval Hard	29.70	28.40	43.30	34.60

Intended Use

The Apriel family of models are designed for a variety of general-purpose instruction tasks, including:

Question answering and information retrieval
Content generation and summarization
Code assistance and generation
Logical reasoning and multi-step tasks
Creative writing and ideation

They are not intended for use in safety-critical applications without human oversight or in scenarios requiring guaranteed factual accuracy.

Limitations

Factual accuracy: May produce incorrect, misleading, or outdated content. Outputs should be verified before use in critical contexts.
Bias: May reflect societal, cultural, or systemic biases present in training data.
Ethics: Do not use the model to produce harmful, unlawful, or unethical content.
Language: Strongest performance is in English. Output quality may degrade in underrepresented languages.
Critical use: Not suitable for medical, legal, financial, or other high-risk applications without safeguards.

Security and Responsible Use

Security Responsibilities:
Deployers and users are strongly encouraged to align their security practices with established frameworks and regulatory guidelines such as the EU AI Act and the NIST AI Risk Management Framework (RMF).

Guidelines for Deployers:

Regularly conduct robustness assessments to identify and mitigate adversarial inputs.
Implement validation and filtering processes to prevent harmful or biased outputs.
Continuously perform data privacy checks to guard against unintended data leaks.
Document and communicate the model's limitations, intended usage, and known security risks to all end-users.
Schedule periodic security reviews and updates to address emerging threats and vulnerabilities.

Guidelines for Users:

Follow established security policies and usage guidelines provided by deployers.
Protect and manage sensitive information when interacting with the model.
Report anomalies, suspicious behavior, or unsafe outputs to deployers or developers.
Maintain human oversight and apply judgment to mitigate potential security or ethical risks during interactions.

Disclaimer:
Users accept responsibility for securely deploying, managing, and using this open-source LLM. The model is provided "as-is," without explicit or implied warranty regarding security or fitness for any specific application or environment.

Pretraining

Model

Architecture: Transformer decoder with grouped-query attention and YARN rotary embeddings
Tokens: 4.5T
Precision: bfloat16
Knowledge cutoff: April 2024

Hardware

Compute: 480 × H100 GPUs
GPU-hours: ~91,000 H100-hours

Software

Training stack: Fast-LLM

License

MIT

Citation

bibtex

@misc{Apriel-small-language-models,  
    author = {Slam labs team},  
    title = {{Apriel - a Family of performant small language models}},  
    howpublished = {https://huggingface.co/ServiceNow-AI/Apriel-5B-Instruct},
    publisher = {SLAM - ServiceNow Language Models Lab}  
    year = {2025}
}

ServiceNow-AI/Apriel-5B-Base

Apriel-5B

Table of Contents

Model Summary

Apriel-5B-Base

Apriel-5B-Instruct

How to Use

Running the Base model

Running the Instruct model

Chat Template

Evaluation

Apriel-5B-Base

Apriel-5B-Instruct

Intended Use

Limitations

Security and Responsible Use

Pretraining

Model

Hardware

Software

License

Citation

ServiceNow-AI/
Apriel-5B-Base