README.md

---
license: apache-2.0
base_model: aquif-ai/aquif-3.6-1B
tags:
- text-generation-inference
- reasoning
- thinking
- hybrid
- efficient
- dynamic
- transformers
- aquif
- math
- coding
- small
- aquif-3.5
- aquif-3.6
- llm
- llama-cpp
- gguf-my-repo
language:
- en
- de
- it
- pt
- fr
- hi
- es
- th
- zh
- ja
library_name: transformers
pipeline_tag: text-generation
---

# 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.

### Contents

- [Key Features](#key-features) - Dynamic reasoning, efficiency gains, and smart resource allocation
- [Performance](#performance) - Benchmark results showing 4% average improvement
- [Token Efficiency](#token-efficiency) - 28% reduction in token usage
- [Thinking Ratio](#thinking-ratio) - 12% reduction in thinking frequency
- [Benchmark Highlights](#benchmark-highlights) - Detailed results for AIME, LiveCodeBench, and GPQA Diamond
- [Model Details](#model-details) - Architecture and specifications
- [Usage](#usage) - Code examples for implementation
- [Previous Versions](#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


## Use with llama.cpp
Install llama.cpp through brew (works on Mac and Linux)

```bash
brew install llama.cpp

```
Invoke the llama.cpp server or the CLI.

### CLI:
```bash
llama-cli --hf-repo Edge-Quant/aquif-3.6-1B-Q4_K_M-GGUF --hf-file aquif-3.6-1b-q4_k_m.gguf -p "The meaning to life and the universe is"
```

### Server:
```bash
llama-server --hf-repo Edge-Quant/aquif-3.6-1B-Q4_K_M-GGUF --hf-file aquif-3.6-1b-q4_k_m.gguf -c 2048
```

Note: You can also use this checkpoint directly through the [usage steps](https://github.com/ggerganov/llama.cpp?tab=readme-ov-file#usage) listed in the Llama.cpp repo as well.

Step 1: Clone llama.cpp from GitHub.
```
git clone https://github.com/ggerganov/llama.cpp
```

Step 2: Move into the llama.cpp folder and build it with `LLAMA_CURL=1` flag along with other hardware-specific flags (for ex: LLAMA_CUDA=1 for Nvidia GPUs on Linux).
```
cd llama.cpp && LLAMA_CURL=1 make
```

Step 3: Run inference through the main binary.
```
./llama-cli --hf-repo Edge-Quant/aquif-3.6-1B-Q4_K_M-GGUF --hf-file aquif-3.6-1b-q4_k_m.gguf -p "The meaning to life and the universe is"
```
or 
```
./llama-server --hf-repo Edge-Quant/aquif-3.6-1B-Q4_K_M-GGUF --hf-file aquif-3.6-1b-q4_k_m.gguf -c 2048
```