Apriel-1.5-Thinker

Table of Contents

1. Summary
2. Evaluation
3. Training Details
4. How to Use
5. Intended Use
6. Limitations
7. Security and Responsible Use
8. Software
9. License
10. Acknowledgements
11. Citation

Click here to skip to the technical report -> https://huggingface.co/ServiceNow-AI/Apriel-1.5-15b-Thinker/blob/main/Apriel-1.5-Thinker.pdf

Summary

Apriel-1.5-15b-Thinker is a multimodal reasoning model in ServiceNow’s Apriel SLM series which achieves competitive performance against models 10 times it's size. Apriel-1.5 is the second model in the reasoning series. It introduces enhanced textual reasoning capabilities and adds image reasoning support to the previous text model. It has undergone extensive continual pretraining across both text and image domains. In terms of post-training this model has undergone text-SFT only. Our research demonstrates that with a strong mid-training regimen, we are able to achive SOTA performance on text and image reasoning tasks without having any image SFT training or RL.

Highlights

• Achieves a score of 52 on the Artificial Analysis index and is competitive with Deepseek R1 0528, Gemini-Flash etc.
• It is AT LEAST 1 / 10 the size of any other model that scores > 50 on the Artificial Analysis index.
• Scores 68 on Tau2 Bench Telecom and 62 on IFBench, which are key benchmarks for the enterprise domain.
• At 15B parameters, the model fits on a single GPU, making it highly memory-efficient.

Evaluation

• For text benchmarks, we report evaluations perforomed by a third party - Artificial Analysis.
• For image benchmarks, we report evaluations obtained by https://github.com/open-compass/VLMEvalKit

Results reported by Artificial Analysis

Note from the developers

We are a small lab with big goals. While we are not GPU poor, our lab, in comparison has a tiny fraction of the compute available to other Frontier labs. Our goal with this work is to show that a SOTA model can be built with limited resources if you have the right data, design and solid methodology.

We set out to build a small but powerful model, aiming for capabilities on par with frontier models. Developing a 15B model with this level of performance requires tradeoffs, so we prioritized getting SOTA-level performance first. Mid-training consists only of CPT and SFT; no RL has been applied.

This model performs extensive reasoning by default, allocating extra internal effort to improve robustness and accuracy even on simpler queries. You may notice slightly higher token usage and longer response times, but we are actively working to make it more efficient and concise in future releases.

Training Details

Training Hardware

• GPUs: 640 × H100
• Training Time: 7 days

Mid training / Continual Pre‑training In this stage, the model is trained on billions of tokens of carefully curated textual samples drawn from mathematical reasoning, coding challenges, scientific discourse, logical puzzles, and diverse knowledge-rich texts along with multimodal samples covering image understanding and reasoning, captioning, and interleaved image-text data. The objective is to strengthen foundational reasoning capabilities of the model. This stage is critical for the model to function as a reasoner and provides significant lifts in reasoning benchmarks.

Supervised Fine‑Tuning (SFT) The model is fine-tuned on over 2M high-quality text samples spanning mathematical and scientific problem-solving, coding tasks, instruction-following, API/function invocation, and conversational use cases. This results in superior text performance comparable to models such as Deepseek R1 0528 and Gemini-Flash. Although no image-specific fine-tuning is performed, the model’s inherent multimodal capabilities and cross-modal transfer of reasoning behavior from the text SFT yield competitive image performance relative to other leading open-source VL models.

Running Apriel-1.5-15B-Thinker with vLLM

As the upstream PR is not yet merged, you can use this custom image as an alternate way to run the model with tool and reasoning parsers enabled.

Docker Image

docker.io/amant555/vllm_apriel:latest

Start Command

Run the container with the following command:

python3 -m vllm.entrypoints.openai.api_server \
  --model ServiceNow-AI/Apriel-1.5-15b-Thinker \
  --served-model-name Apriel-1p5-15B-Thinker \
  --trust_remote_code \
  --max-model-len 131072 \
  --enable-auto-tool-choice \
  --tool-call-parser apriel \
  --reasoning-parser apriel

This will start the vLLM OpenAI-compatible API server serving the Apriel-1.5-15B-Thinker model with Apriel’s custom tool parser and reasoning parser.

How to Use

pip install transformers

Running the Reasoning model

Here is a code snippet demonstrating the model's usage with the transformers library's generate function:

# Tested with transformers==4.48

import re
import requests
import torch
from PIL import Image
from transformers import AutoProcessor, AutoModelForImageTextToText

# Load model
model_id = "ServiceNow-AI/Apriel-1.5-15b-Thinker"
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16,
    device_map="auto"
)
processor = AutoProcessor.from_pretrained(model_id)

# Example 1: Text-only prompt
chat = [
    {
        "role": "user",
        "content": [
            {"type": "text", "text": "What is the capital for France?"},
        ],
    }
]

inputs = processor.apply_chat_template(chat, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt")
inputs = {k: v.to(model.device) if isinstance(v, torch.Tensor) else v for k, v in inputs.items()}
inputs.pop("t