aquif-3.0 Preview 2

aquif-3.0 Preview 2 is the second iteration of the aquif-3.0 model, bringing major advancements in reasoning and accuracy. This release introduces thinking mode, a novel capability that unlocks up to 10% overall improvement by simply sending a control message, along with a bigger 128 thousand token context window, up from 16,364.

Model Overview

• Name: aquif-3.0-preview-2
• Parameters: 2.5 Billion
  
• Context Window: 128,000 tokens
  
• Architecture: Decoder-only transformer
  
• Type: General-purpose LLM
  
• Hosted on: Ollama

Features

• 128K token context for extended reasoning and document analysis
  
• Optional thinking mode for boosted reasoning (via control message)
  
• Strong performance in math, logic, and code benchmarks
  
• Offline-friendly, optimized for fast inference on local machines
  

Activating Thinking Mode

To enable enhanced reasoning, send the following message before your prompt:

{
  "role": "control",
  "content": "thinking"
}

This activates a mode that performs deeper internal chain-of-thought reasoning.

Performance Benchmarks

aquif-3.0 Preview 2 shows strong gains over Preview 1, and the preview-2-thinking variant outperforms other open-weight 8B models across multiple benchmarks:

*These are estimated results. The model’s final version is releasing later this year.

Use Cases

• Chain-of-thought tasks and multi-step reasoning
  
• Code generation and evaluation
  
• Document question answering and long-context workflows
  
• Planning, agents, and assistant-style applications

Limitations

• The model often hallucinates on knowledge questions
• Unlike preview-1, this version still does not include vision capabilities

Getting Started

To run via Ollama:

ollama run aquiffoo/aquif-3.0-preview-2

To enable thinking mode:

ollama run aquiffoo/aquif-3.0-preview-2

Then, send this message before any user prompt:

{ "role": "control", "content": "thinking" }