KL/D and Refusals statistic on Heretic cannot be an independent measurement what's true or what best uncensored means. As far i can tell, consistently hereticsation with one model; picking high refusal score with low KL/D value sometimes good, sometimes bad. It's inconsistent (unpredictable) and vice versa. For example, higher refusals count (e.g. 10/100) from creator "Z" can be better than "A" who has 0/100 refusals. We never truly know the 'refusal marker recipe' other creators use just to make their scores look better or to reach a 0/100 goal.
Independent measurement, benchmarking, your own subjective judgement carry more weight than this shallow metric.
If you feel heretic model is not good for your tastes, give us a feedback, don't be Hidey Harold. Report.
Apriel-1.6-15B-Thinker is an updated multimodal reasoning model in ServiceNow’s Apriel SLM series, building on Apriel-1.5-15B-Thinker.
With significantly improved text and image reasoning capabilities, Apriel-1.6 achieves competitive performance against models up to 10x its size.
Like its predecessor, it benefits from extensive continual pre-training across both text and image domains.
We additionally perform post-training that focuses on Supervised Finetuning (SFT) and Reinforcement Learning (RL).
Apriel-1.6 obtains frontier performance without sacrificing reasoning token efficiency.
The model improves or maintains task performance when compared with Apriel-1.5-15B-Thinker, while reducing reasoning token usage by more than 30%.
Highlights
Achieves a score of 57 on the Artificial Analysis index outperforming models like Gemini 2.5 Flash, Claude Haiku 4.5 and GPT OSS 20b. It obtains a score on par with Qwen3 235B A22B, while being significantly more efficient.
Reduces reasoning token usage by more than 30%, delivering significantly better efficiency than Apriel-1.5-15B-Thinker.
Scores 69 on Tau2 Bench Telecom and 69 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.
Based on community feedback on Apriel-1.5-15B-Thinker, we simplified the chat template by removing redundant tags and introduced four special tokens to the tokenizer (<tool_calls>, </tool_calls>, [BEGIN FINAL RESPONSE], <|end|>) for easier output parsing.
Text benchmarks included in the Artificial Analysis Index v3.0 use scores reported by Artificial Analysis. All other benchmarks were evaluated internally.
Category
Benchmark
Apriel-1.6-15B-Thinker
Apriel-1.5-15B-Thinker
GPT OSS 120B
DeepSeek R1 0528
Gemini 2.5 Flash (Sep)
GPT 5 mini (high)
Claude 4.5 Sonnet (thinking)
o3-mini (high)
Average Score**
53.62
46.56
52.56
51.92
50.71
62.58
60.37
48.85
Function Calling
BFCL v3 only
63.50
51.88
50.62
39.75
39.75
17.62
-
50
Tau2 bench Telecom
69
57.8
66
37
32
68
50.8
31
Tau2 bench Retail
66.67
46.78
61.4
59.94
61.69
73.39
69.8
75.73
Tau2 bench Airline
58
52
45.3
47.33
56.66
59.33
58
61.33
ComplexFuncBench
33.2
19
24.6
24.2
26.3
37.5
24.6
18.9
Instruction Following
Agent IF
57.2
55
54.20
52.20
49.70
57.60
54.50
54.90
Multi IF
83.34
76.91
82.95
73.76
82.49
85.37
84.32
87.28
Multi-Challenge
46.15
41.39
46.90
44.50
49.08
57.90
42.49
38.46
IF Bench
69
62
69
40
50
75
57
70.07
Math
AIME 25
88
88
93
76
73
91
88
86.67
Coding
Struct Eval
79
48.50
71
73
70
69.92
76
73
LCB
81
73
88
77
70
84
71
73
SciCode
37
35
39
40
41
39
45
40
Agentic
DeepresearchBench
36.47
32.73
36.30
34.19
38.15
-
-
33.40
GAIA
40
30.91
21.21
32.12
47.88
65.45
69.09
23.03
Work-Arena L1
59.1
51.5
50.9
63.9
51.8
65.5
62.7
52.4
OS World Small
16.70
13.90
16.70
25
19.40
22.20
30.60
19.40
SWE Bench Verified
23
16
31
29.60
34.20
61
64.2
22.60
Terminal Bench
14
10
22
15
13
31
33
5.67
Aider Polyglot
37.68
26.37
42
71.40
40
71.60
78
60.40
Knowledge
MMLU Pro
79
77
81
85
83
84
88
80
Creative Writing
Creative writing v3 / EQ Bench
59.73
60.24
53.70
79.40
74.25
75.25
80.70
30.40
Others
GPQA Diamond
73
71
78
81
79
83
83
77
HLE
10
12
18.5
14.9
11.1
19.7
17.3
12.3
Long Context
AA LCR
50*
20
51
55
62
68
66
30***
* This score is with DCA enabled. Without this, the model scores 36.
** The average score is calculated using all benchmarks except BFCL v3 Only and DeepResearchBench, since some models do not have scores for these two benchmarks.
*** AA LCR score for o3-mini-high is projected score based on its AA Index score.
The Apriel family of models are designed for a variety of general-purpose instruction tasks, including:
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.
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:
python
1# Tested with transformers==4.4823import re
4import requests
5import torch
6from PIL import Image
7from transformers import AutoProcessor, AutoModelForImageTextToText
89# Load model10model_id ="ServiceNow-AI/Apriel-1.6-15b-Thinker"11model = AutoModelForImageTextToText.from_pretrained(12 model_id,13 torch_dtype=torch.bfloat16,14 device_map="auto"15)16processor = AutoProcessor.from_pretrained(model_id)1718# Example 1: Text-only prompt19chat =[20{21"role":"user",22"content":[23{"type":"text","text":"What is the capital for France?"},24],25}26]2728inputs = processor.apply_chat_template(chat, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors="pt")29inputs ={k: v.to(model.device)ifisinstance(v, torch.Tensor)else v for k, v in inputs.items()}30inputs.pop("token_type_ids",None)3132with torch.no_grad():33 output_ids = model.generate(**inputs, max_new_tokens=1024, do_sample=True, temperature=0.6)3435generated_ids = output_ids[:, inputs['input_ids'].shape[1]:]36output = processor.decode(generated_ids[0], skip_special_tokens=True)37response = re.findall(r"\[BEGIN FINAL RESPONSE\](.*?)(?:<\|end\|>)", output, re.DOTALL)[0].strip()3839print("Text-only Response:", response)4041# Example 2: Image understanding42url ="https://picsum.photos/id/237/200/300"43image = Image.open(requests.get(url, stream=True).raw).convert("RGB")4445chat =[46{47"role":"user",48"content":[49{"type":"text","text":"Which animal is this?"},50{"type":"image"},51],52}53]5455prompt = processor.apply_chat_template(chat, add_generation_prompt=True, tokenize=False)56inputs = processor(text=prompt, images=[image], return_tensors="pt").to(model.device)5758with torch.no_grad():59 output_ids = model.generate(**inputs, max_new_tokens=1024, do_sample=True, temperature=0.6)6061generated_ids = output_ids[:, inputs['input_ids'].shape[1]:]62output = processor.decode(generated_ids[0], skip_special_tokens=True)63response = re.findall(r"\[BEGIN FINAL RESPONSE\](.*?)(?:<\|end\|>)", output, re.DOTALL)[0].strip()6465print("Image Response:", response)66
Usage Guidelines
Use the model’s default chat template, which already includes a system prompt.
We recommend setting temperature to 0.6.
We ensure the model starts with Here are my reasoning steps:\n during all our evaluations. This is implemented in the default chat template.
For multi-turn conversations, intermediate turns (historical model outputs) are expected to contain only the final response, without reasoning steps.
Chat Template
<|begin_system|>
You are a thoughtful, systematic AI assistant from ServiceNow Language Models (SLAM) lab. Analyze each question carefully, present your reasoning step-by-step, then provide the final response after the marker [BEGIN FINAL RESPONSE].
<|begin_user|>
# user message here
<|begin_assistant|>
Here are my reasoning steps:
# thoughts here
[BEGIN FINAL RESPONSE]
# assistant response here
<|end|>
The model will first generate its thinking process and then generate its final response, starting with [BEGIN FINAL RESPONSE]. Here is a code snippet demonstrating the application of the chat template:
python
1from transformers import AutoTokenizer
2model_name ="ServiceNow-AI/Apriel-1.6-15b-Thinker"3tokenizer = AutoTokenizer.from_pretrained(model_name)45# prepare the model input6custom_system_prompt ="Answer like a pirate."7prompt ="You are an expert assistant in the implementation of customer experience management aspect of retail applications \n \nYou will be using Python as the programming language. \n \nYou will utilize a factory design pattern for the implementation and following the dependency inversion principle \n \nYou will modify the implementation based on user requirements. \n \nUpon user request, you will add, update, and remove the features & enhancements in the implementation provided by you. \n \nYou will ask whether the user wants to refactor the provided code or needs a sample implementation for reference. Upon user confirmation, I will proceed accordingly. \n \n**Guidelines:** \n 1. **User Requirements:** \n - You have to ask users about their requirements, clarify the user expectations, and suggest the best possible solution by providing examples of Python code snippets. \n - Ask users about which type of reports they need to assess the AI model's performance, accuracy, and reliability. \n - After providing the solution, you have to ask the user about the trial of the solution and modify the solution based on the user feedback. \n \n 2. **Libraries/Frameworks:** \n - You will be utilizing Python as a programming language. \n - You will be using Flask framework for REST APIS implementation \n \n 3. **Communication Gesture:** \n - Your conversation with the user should be interactive, supportive, courageous, and professional. \n - You have to break down the complex concepts into sub-concepts and try to explain them to the user. \n - You have to ask the user for the required parameters. If the user refuses to provide in 2 attempts, politely exit the conversation. \n - You have to provide your supported parameters to the user, if the user refuses to accept them then you have to put an apology note and exit the conversation. \n - You have to track the conversation about unasked questions by the user. If some/one of the questions remain then you have to remind the user about these questions and proceed to answer them based on the user's confirmation \n \n 4. **Implementation:** \n - Your code/implementations should be reliable, scaleable, modular, and reusable. \n - You will be providing unit tests for the implementation upon user request. \n - You will be following MVC architecture for the applications \n - Your implementations must be well-commented and readable \n \n \n- Today's date is 23rd August 2024. \n- The default sender email is sender-assistant@email.com.\nHi, I am conducting research on retail customer feedback systems and I need assistance with designing and implementing them. Could you kindly provide me with a list of general customer feedback system modules?"8messages =[9{"role":"user","content": custom_system_prompt +"\n\n"+ prompt}10]11# example tools12tools =[{"type":"function","function":{"name":"getRetailFeedbackModules","description":"Returns the list of modules usually present in the retail industry","parameters":{"type":"object","properties":{"page":{"type":"integer","description":"The current page number.","default":1},"page_size":{"type":"integer","description":"The number of items per page.","default":3}}}}},{"type":"function","function":{"name":"verifyImplementation","description":"Returns the list of modules usually present in the retail industry","parameters":{"type":"object","properties":{"coding_language":{"type":"string","description":"The supported languages for verification of implementation.","default":"python","enum":["python","java","php"]},"code":{"type":"string","description":"The code which needs verification"},"design_pattern":{"type":"string","description":"The design pattern to verify in the implementation","enum":["factory","strategy","singleton"]},"verify_best_practices":{"type":"boolean","description":"The verification of the coding style based on the language selected","default": true}}}}}]13text = tokenizer.apply_chat_template(14 messages,15 tokenize=False,16 add_generation_prompt=True,17 tools=tools
18)19model_inputs = tokenizer([text], return_tensors="pt")
Running 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.
Continual Pre-training: Billions of tokens covering math, code, science, logical reasoning, and multimodal image-text data
SFT: 2.4M samples spanning math, code, instruction-following, function calling, and conversation, followed by an incremental lightweight multimodal SFT.
RL: Multi-stage RL with verifiable rewards and GSPO on text and vision tasks. Our RL stage optimizes reasoning efficiency: using fewer tokens by discouraging unnecessary intermediate steps, stopping earlier when confident, and giving direct answers on simple queries.
For more details on our training methodology, see our blog post.
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.
License
MIT
Citation
bibtex
1@misc{radhakrishna2025apriel1515bthinker,
2 title={Apriel-1.5-15b-Thinker},
3 author={Shruthan Radhakrishna and Aman Tiwari and Aanjaneya Shukla and Masoud Hashemi and Rishabh Maheshwary and Shiva Krishna Reddy Malay and Jash Mehta and Pulkit Pattnaik and Saloni Mittal and Khalil Slimi and Kelechi Ogueji and Akintunde Oladipo and Soham Parikh and Oluwanifemi Bamgbose and Toby Liang and Ahmed Masry and Khyati Mahajan and Sai Rajeswar Mudumba and Vikas Yadav and Sathwik Tejaswi Madhusudhan and Torsten Scholak and Sagar Davasam and Srinivas Sunkara and Nicholas Chapados},
4 year={2025},
5 eprint={2510.01141},
6 archivePrefix={arXiv},
7 primaryClass={cs.AI},
8 url={https://arxiv.org/abs/2510.01141},
9}
