aquif-3.6-1B

Summary

aquif-3.6-1B is a hybrid reasoning model that automatically determines when and how deeply to think based on query complexity. Built on aquif-3.5-Nano-1B with AutoThink RL data, it achieves 28% better token efficiency and 4% performance improvement across benchmarks.

Key Features - Dynamic reasoning, efficiency gains, and smart resource allocation
Performance - Benchmark results showing 4% average improvement
Token Efficiency - 28% reduction in token usage
Thinking Ratio - 12% reduction in thinking frequency
Benchmark Highlights - Detailed results for AIME, LiveCodeBench, and GPQA Diamond
Model Details - Architecture and specifications
Usage - Code examples for implementation
Previous Versions - Links to earlier models

Automatic Thinking

aquif-3.6-1B is a hybrid reasoning model that dynamically decides if and how much to think based on query complexity. Inspired by aquif-3.6-8B's approach of automatic thinking using AutoThink RL data on top of aquif-3.5-Nano-1B, the model uses the following format:

<judge>
[analyzes whether to think or not]
</judge>

<think_on/off>
<think>
[thinking content]
</think>

<answer>
</answer>

This is the same format as aquif-3.6-8B. Unlike something like aquif-3.5-Plus's toggleable reasoning that requires manual control (thinking_on/off), aquif-3.6's judge autonomously allocates reasoning depth - intelligently adapting its cognitive effort to each task automatically.

Key Features

🧠 Dynamic Reasoning: Automatically determines when and how deeply to think
⚡ 28% More Efficient: Significant token reduction while improving performance
📈 Better Performance: 4% average improvement across benchmarks
🎯 Smart Resource Allocation: 12% reduction in thinking ratio on average

Performance

Benchmark	aquif-3.6-1B	Qwen3-1.7B	Improvement
AIME 2025	75.0	39.4	+35.6%
LiveCodeBench	57.5	33.2	+24.3%
GPQA Diamond	52.8	40.1	+12.7%
Average	61.8	37.6	+24.2%

Token Efficiency

Benchmark	aquif-3.6-1B	Qwen3-1.7B	Reduction
AIME 2025	13,670	18,450	-26%
LiveCodeBench	10,270	13,890	-26%
GPQA Diamond	6,870	12,100	-43%
Average	10,270	14,813	-32%

Thinking Ratio

Benchmark	aquif-3.6-1B	Qwen3-1.7B	Reduction
AIME 2025	84.0%	100.0%	-16%
LiveCodeBench	78.0%	100.0%	-22%
GPQA Diamond	81.0%	100.0%	-19%
Average	81.0%	100.0%	-19%

Benchmark Highlights

AIME 2025: 26% fewer tokens, +35.6% performance, -16% thinking ratio
LiveCodeBench: 26% fewer tokens, +24.3% performance, -22% thinking ratio
GPQA Diamond: 43% fewer tokens, +12.7% performance, -19% thinking ratio

Model Details

Base Model: 1.7B parameters
Architecture: Hybrid reasoning with dynamic thinking allocation
Context Length: 40K tokens
License: Apache 2.0

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "aquif-ai/aquif-3.6-1B"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype="auto",
    device_map="auto"
)

messages = [
    {"role": "user", "content": "Solve this problem: What is the sum of all prime numbers between 1 and 100?"}
]

input_ids = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt"
).to(model.device)

outputs = model.generate(
    input_ids,
    max_new_tokens=2048,
    temperature=0.7,
    do_sample=True
)

response = tokenizer.decode(outputs[0][input_ids.shape[-1]:], skip_special_tokens=True)
print(response)