ServiceNow Expands Its AI Ambitions With New Apriel Small Language Models

ServiceNow, the enterprise automation platform, has released a family of small language models called Apriel, available in two variants: a 5B base model and an instruct-tuned version. The company lists that they are designed to handle general-purpose tasks, from answering questions and retrieving information, to content generation, code assistance, reasoning, and creative writing. Both models are now available for download from Hugging Face.

“Apriel is a small, high-performing language model ServiceNow built and trained entirely in-house, from data to architecture to infrastructure. It delivers strong results while using far less compute than comparable models, and helps us explore how to build smarter, more efficient AI,” Torsten Scholak, ServiceNow’s research lead at its Foundation Models Lab, tells me in a statement.

Named after what appears to be a variant of the word “April,” and rooted in the Latin word aperire, meaning “to open,” the Apriel model family is the first to emerge from ServiceNow’s Language Modeling Lab. The timing of the release in April, along with the knowledge cutoff date of April 2024, suggests this naming is no coincidence. However, it is interesting that the cutoff date is not more recent, considering the rapid pace of language model development.

While it’s not the company’s first foray into open-source AI—ServiceNow partnered with Hugging Face in 2023 on the open-access StarCoder 15B model—it’s rare to see a fully home-grown release.

Apriel-5B-Base vs. Instruct

Apriel-5B-Base is a decoder-only model trained on more than 4.5 trillion tokens of data. It’s intended to be fine-tuned or adapted for specific tasks as part of a broader AI workflow. ServiceNow claimed it performs better than similar 5B parameter models across common benchmarks. Because it has such a small parameter, Apriel-5B-Base is likely intended for edge devices such as tablets, phones, and wearables.

The second variation, Apriel-5B-Instruct, builds on the base model with additional training that makes it suitable for answering questions, following prompts, and having safe and helpful conversations. The company states Apriel-5B-Instruct was refined using continual pretraining, supervised fine-tuning, and post-training alignment techniques including Direct Preference Optimization (DPO) and Reinforcement Learning with Verifiable Rewards (RLVR).

Although ServiceNow points out its Apriel family of models can be used for a variety of general-purpose assignments, it warns that it shouldn’t be used for “safety-critical applications” in the absence of human oversight or “scenarios requiring guaranteed factual accuracy.”

How Apriel Stacks Up to Competing Models

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

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

Why Does ServiceNow Need an SLM?

ServiceNow’s decision to develop its own small language model, Apriel, likely stems from the company’s desire to have more control and customization over an AI assistant or application that can integrate seamlessly with its enterprise automation platform. By creating an SLM tailored to its specific needs, ServiceNow can better optimize the model’s performance for task-oriented applications within its ecosystem, especially as the company expands into the CRM space. This move also allows the company to explore innovative techniques in responsible AI development and push the boundaries of what smaller language models can achieve.

Apriel’s release is likely part of a broader strategic push by ServiceNow to cement its position as a leader in enterprise AI. The company has made significant investments in this space, most notably the $2.9 billion acquisition of the agentic AI platform Moveworks and launching AI agents on Now Assist and Now Platform. By developing and releasing its own small language model, ServiceNow is sending a clear signal about its AI capabilities and ambitions.

This move allows ServiceNow to showcase its technical expertise and commitment to driving innovation in the enterprise automation market. The Apriel models can be seamlessly integrated into the company’s core platform, enabling a new generation of AI-powered applications and services tailored to the needs of its customer base.

“While Apriel-5B is not a ServiceNow product and is not deployed in any customer environment, this research helps inform ServiceNow’s broader work in enterprise AI,” Scholak says. “Insights from this work are being used to improve model efficiency and performance in future initiatives.”

Beyond just showcasing its AI prowess, ServiceNow’s investment in Apriel underscores the company’s vision for the future of work, where intelligent automation and natural language processing play a central role. By developing its own small language model, ServiceNow is positioning itself to stay ahead of the curve and deliver cutting-edge AI solutions that drive productivity, efficiency, and customer satisfaction for its enterprise clients.

Scholak adds, “Apriel-5B is an open-weight model released by researchers at ServiceNow as part of their ongoing work on efficient language model training. It reflects ServiceNow’s commitment to contributing to the open-source ecosystem and advancing the science behind smaller, more efficient models.”

You can download Apriel-5B-Base and Apriel-5B-Instruct on Hugging Face today.

Updated on April 21, 2025: Included statements from ServiceNow

Featured Image: An AI-generated image of a compact neural network with the ServiceNow logo overlaid on top. Image credit: Adobe Firefly