ServiceNow-AI / Apriel-Reasoner-v1

Apriel-Reasoner

Apriel-Reasoner is a 15B open-weight reasoning model trained with reproducible multi-domain RL post-training using PipelineRL on top of Apriel-1.5-15b-Thinker . It is designed to improve reasoning quality under practical output-token budgets across mathematics, code generation, instruction following, logical puzzles, and function calling.

Model Details

• Model name: Apriel-Reasoner
• Base model: ServiceNow-AI/Apriel-1.5-15b-Thinker
• Training stack: PipelineRL
• Parameters: 15B
• Model type: causal language model
• Language: English
• Training regime: multi-domain RL with verifiable rewards (RLVR)
• Training output budget: 16K output tokens
• Evaluated output budget: up to 32K output tokens

Training

Apriel-Reasoner was trained from Apriel-1.5-15b-Thinker using PipelineRL for asynchronous multi-domain RL post-training. Key ingredients include:

• PipelineRL as the distributed RL training stack
• Adaptive domain sampling to preserve target domain ratios under asynchronous rollout dynamics
• Difficulty-Aware Length Penalty (DAP) to encourage concise reasoning on easier problems while allowing longer reasoning on harder ones

The training recipe uses RL with verifiable rewards across five domains:

• Mathematics
• Code generation
• Instruction following
• Logical puzzles
• Function calling

Selected training mixture:

• Mathematics: 40%
• Code generation: 25%
• Logical puzzles: 15%
• Instruction following: 10%
• Function calling: 10%

Intended Use

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

• Mathematical reasoning
• Code assistance and generation
• Logical reasoning and multi-step tasks
• Question answering and information retrieval
• Function calling, complex instruction following and agent use cases 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• Factual accuracy: May produce incorrect, misleading, or outdated content.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.

Evaluation

Apriel-Reasoner was evaluated on:

• AIME 2025
• GPQA
• MMLU-Pro
• LiveCodeBench Reported highlights from the paper:
• Improves over Apriel-1.5-15b-Thinker on AIME 2025, GPQA, MMLU-Pro, and LiveCodeBench
• Produces 30-50% shorter reasoning traces
• Generalizes to a 32K output budget despite being trained with a 16K output budget

Usage

Use the same tokenizer and prompting conventions as Apriel-1.5-15b-Thinker.

Transformers example

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch
model_name = "ServiceNow-AI/Apriel-Reasoner-v1"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.bfloat16,
    device_map="auto",
)
prompt = "Solve the following problem carefully:\n\nIf x^2 - 5x + 6 = 0, what are the solutions?"
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    outputs = model.generate(
        **inputs,
        max_new_tokens=2048,
        do_sample=True,
        temperature=0.6,
        top_p=0.95,
    )
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Data

Training uses public datasets and verifiable reward functions across five domains, including:

• Open-Reasoner-Zero for mathematics
• TACO for code generation
• IF-RLVR for instruction following
• INTELLECT-3 / SynLogic for logical puzzles
• BFCL v4 single-turn tasks for function calling See the paper for full dataset and verifier details.

Software

• Training stack: PipelineRL

License

This model is released under the MIT License. Please separately check the licenses and terms of:

• the base model
• the training datasets
• the evaluation datasets
• any downstream application or deployment setting

Citation

Apriel-Reasoner paper link and citation will be added here once available.
@article{apriel_reasoner_2026,
  title   = {Apriel-Reasoner: RL Post-Training for General-Purpose and Efficient Reasoning},
  author  = {Rafael Pardinas and Ehsan Kamalloo and David Vazquez and Alexandre Drouin},
  year    = {2026},
  journal = {arXiv preprint},
  note    = {placeholder citation}
}

Links

• Apriel-Reasoner paper: coming soon
• Training pipeline / code: PipelineRL
• Related paper: PipelineRL: Faster On-policy Reinforcement Learning for Long Sequence Generation
• Base model: Apriel-1.5-15b-Thinker