Apriel-1.5-15b-Thinker

Frontier-level multimodal reasoning in a compact, efficient model

Try Now

read docs

About model

Apriel-1.5-15b-Thinker is a breakthrough multimodal reasoning model from ServiceNow's Apriel SLM series that achieves frontier-level performance despite being just 15 billion parameters. Built through innovative mid-training techniques, this model shows that thoughtful data curation and staged continual pretraining can rival systems 10x its size.

AA Intelligence Index

Competitive with DeepSeek-R1 & Gemini-2.5-Flash

AIME'25 Accuracy

87%

Elite mathematical reasoning

Deployment Footprint

1 GPU

Single-GPU efficiency at 15B params

Model key capabilities

Mathematical Reasoning: 87% on AIME'25
Enterprise Benchmarks: 68% Tau2 Bench Telecom, 62% IFBench
Multimodal Understanding: Text and image reasoning across domains
Accessible Deployment: Frontier AI on a single GPU

Quickstart guides

Building a RAG Workflow

API usage

cURL
Python
Typescript

Endpoint:

ServiceNow-AI/Apriel-1.5-15b-Thinker

curl -X POST "https://api.together.xyz/v1/chat/completions" \
  -H "Authorization: Bearer $TOGETHER_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "ServiceNow-AI/Apriel-1.5-15b-Thinker",
    "messages": [
      {
        "role": "user",
        "content": "What are some fun things to do in New York?"
      }
    ]
}'

from together import Together

client = Together()

response = client.chat.completions.create(
  model="ServiceNow-AI/Apriel-1.5-15b-Thinker",
  messages=[
    {
      "role": "user",
      "content": "What are some fun things to do in New York?"
    }
  ]
)
print(response.choices[0].message.content)

import Together from 'together-ai';
const together = new Together();

const completion = await together.chat.completions.create({
  model: 'ServiceNow-AI/Apriel-1.5-15b-Thinker',
  messages: [
    {
      role: 'user',
      content: 'What are some fun things to do in New York?'
     }
  ],
});

console.log(completion.choices[0].message.content);

Model card
Architecture & Foundation:
• Built from Pixtral-12B base using depth upscaling (40 to 48 layers) for enhanced reasoning capacity
• Multimodal architecture with vision encoder, projection network, and decoder supporting both text and image inputs
• 14.9B parameters optimized for single-GPU deployment with BF16 precision
• 131K token context window with sequence packing for efficient processing

Training Methodology:
• Three-stage progressive training: depth upscaling, staged continual pretraining (CPT), and supervised fine-tuning (SFT)
• CPT Stage 1: 50% text reasoning, 20% replay data, 30% multimodal tokens covering documents, charts, OCR, and visual reasoning
• CPT Stage 2: Targeted visual reasoning via synthetic data generation for spatial structure, compositional understanding, and fine-grained perception
• Text-SFT only approach with 2M+ high-quality instruction-response pairs featuring explicit reasoning traces—no reinforcement learning or preference optimization
• Trained on 640 H100 GPUs for 7 days using Fast-LLM training stack

Performance Characteristics:
• Achieves 52 on Artificial Analysis Intelligence Index, matching DeepSeek-R1-0528 and Gemini-2.5-Flash
• Strong mathematical reasoning: 87% AIME'25, 77.3% MMLU-Pro, 71.3% GPQA Diamond
• Enterprise-focused benchmarks: 68% Tau2 Bench Telecom, 62% IFBench
• Multimodal capabilities: 70.2% MMMU, 75.5% MathVista, 88.2% CharXiv descriptive, 82.87% AI2D
• Extensive reasoning by default with explicit thinking steps before final responses
• Performs within 5 points of Gemini-2.5-Flash and Claude Sonnet-3.7 across ten vision benchmarks
• At least 1/10 the size of any model scoring >50 on AA Intelligence Index
‍
Applications & use cases
Mathematical & Scientific Reasoning:
• Competition-level mathematics: 87% on AIME'25, 80.66% on AIME'24
• Graduate-level problem solving: 71.3% on GPQA Diamond
• Scientific computing and reasoning tasks with strong performance on SciCode
• Mathematical reasoning within visual contexts (MathVision, MathVista, MathVerse)

Code Assistance & Development:
• Functional correctness in code generation via LiveCodeBench evaluation
• Coding tasks spanning multiple programming languages
• API/function invocation and complex instruction following
• Real-world Linux shell execution and system tool use (TerminalBench)

Enterprise & Domain-Specific Applications:
• Specialized telecom domain tasks: 68% on Tau2 Bench Telecom
• Instruction following and compliance: 62% on IFBench
• Document understanding, chart interpretation, and OCR-related tasks
• Long-context reasoning (AA-LCR benchmark) for extended document analysis

Multimodal Understanding:
• Image understanding and reasoning: 70.2% MMMU, 66.3% MMStar
• Document and diagram comprehension: 88.2% CharXiv descriptive, 82.87% AI2D
• Visual mathematical problem-solving: 75.5% MathVista
• Chart understanding with descriptive and reasoning capabilities

General-Purpose Capabilities:
• Multi-domain knowledge and advanced reasoning (77.3% MMLU-Pro)
• Conversational AI and question answering across diverse topics
• Logical reasoning and multi-step task execution
• Content moderation, security, and robustness applications
• On-premises deployment for privacy-sensitive and air-gapped environments
‍

Related models

Model specifications

Model data

Model provider
ServiceNow AI
Type
Reasoning
Vision
Chat
Main use cases
Chat
Vision
Reasoning
Deployment
Serverless
Endpoint
ServiceNow-AI/Apriel-1.5-15b-Thinker
Parameters
15B
Context length
128K
Input modalities
Text
Image
Output modalities
Text

Released
September 23, 2025
Quantization level
BF16
External link
Provider docs
Category
Chat

Run in Playground

Quickstart docs

Deploy model

Apriel-1.5-15b-Thinker

About model

API usage

Model card

Applications & use cases