Running Gemini 2.0 Models on macOS
Published on February 8, 2025   •   6 min read

Running Gemini 2.0 Models on macOS

Jarek CeborskiJarek Ceborski

Google has just released their next generation of AI models - the Gemini 2.0 family. This major update brings significant improvements in performance, capabilities, and cost-effectiveness. Let's explore what these new models offer and how you can easily use them on your Mac with Kerlig.

The Gemini 2.0 Family

Gemini 2.0 Flash

The workhorse of the family, Flash is designed for high-volume, production-ready tasks. It offers:

  • 1 million token context window
  • Multimodal capabilities (text, image, audio, video input)
  • Native tool use
  • Excellent performance/cost ratio
  • Best suited for: Chat applications, content generation, and document analysis

Gemini 2.0 Flash-Lite

A new cost-optimized model that maintains quality while being incredibly efficient:

  • Same 1 million token context window as Flash
  • Multimodal input support
  • 75% lower cost than Flash
  • Best suited for: High-scale applications where cost optimization is crucial

Gemini 2.0 Pro

The most capable model in the family:

  • 2 million token context window (largest available)
  • Superior coding performance
  • Enhanced reasoning capabilities
  • Best suited for: Complex coding tasks and advanced reasoning

Gemini 2.0 Flash Thinking (Experimental)

A specialized model that reveals its reasoning process:

  • Enhanced performance on math and science tasks
  • Same 1M token context window as Flash
  • Shows its step-by-step thinking process
  • Best suited for: Complex problem-solving where explainability is crucial

The Flash Thinking model demonstrates remarkable improvements in specific areas:

  • 73.3% accuracy on AIME2024 math problems (compared to 35.5% for standard Flash)
  • 74.2% on GPQA Diamond science questions (compared to 58.6% for standard Flash)
  • 75.4% on multimodal reasoning tasks (MMMU benchmark)

What makes this model unique is its ability to show its reasoning process, making it especially valuable for:

  • Complex mathematical calculations
  • Scientific problem-solving
  • Educational applications where understanding the thought process is important
  • Tasks requiring transparent decision-making

Model Comparison

ModelContext WindowInput Price*Output Price*Best For
Gemini 2.0 Flash1M tokens$0.10$0.40General purpose, production
Gemini 2.0 Flash-Lite1M tokens$0.075$0.30Cost-efficient scaling
Gemini 2.0 Pro2M tokens--Complex reasoning & coding
Gemini 2.0 Flash Thinking1M tokens--Explainable problem-solving

*Price per 1M tokens

Benchmark Results

The Gemini 2.0 family shows impressive performance across various benchmarks:

General Performance

  • Pro leads with 79.1% on general MMLU-Pro tests
  • Flash achieves 77.6% on diverse subject matter tasks
  • Even the cost-effective Flash-Lite maintains strong 71.6% performance

Specialized Capabilities

  • Mathematics: Pro excels at complex math with 91.8% on MATH benchmark
  • Coding: Up to 36% accuracy on LiveCodeBench for Python generation
  • Reasoning: Strong performance on GPQA Diamond (up to 64.7% for Pro)
  • Multilingual: Excellent language capabilities with Pro reaching 86.5% on Global MMLU

Multimodal Skills

  • Strong video analysis (71.9% on EgoSchema)
  • Capable image understanding (72.7% on MMMU)
  • Competitive audio processing (40.6% on CoVoST2)

These benchmarks demonstrate that while Pro offers the highest performance, both Flash and Flash-Lite maintain strong capabilities across most tasks, making them excellent choices for production use cases.

Community Evaluation

According to LMArena.ai↗, a community-driven AI model evaluation platform with over 2.6 million user votes, Gemini 2.0 models are currently leading the leaderboard:

LMArena.ai Leaderboard
  • Gemini 2.0 Flash Thinking ranks #1 with an Arena Score of 1383
  • Gemini 2.0 Pro follows closely at #2 with a score of 1378
  • Both models outperform competitors like ChatGPT-4 (1365) and DeepSeek R1 (1362)

This real-world, community-driven evaluation demonstrates that Gemini 2.0 models not only excel in controlled benchmarks but also deliver superior performance in practical, day-to-day use cases.

Running Gemini 2.0 on Your Mac

Kerlig makes it easy to access these powerful models right from your macOS desktop. Here's how to get started:

1. Get Your API Key

First, you'll need to get your Google AI API key:

  1. Visit Google AI Studio↗
  2. Log in with your Google account
  3. Click "Get API Key" and copy it
  4. Enter the key in Kerlig's settings

2. Choose Your Model

Kerlig Settings Google Gemini Models

In Kerlig, you can select different Gemini models based on your needs:

  • Use Flash for everyday tasks and general content creation
  • Choose Flash-Lite when working with high volumes of content
  • Select Pro for complex coding tasks or when you need advanced reasoning

3. Start Creating

With Kerlig's native macOS interface, you can:

  • Use ⌥ Option + Space to quickly access AI anywhere
  • Process multiple file types including PDFs, images, and code
  • Get AI assistance in any application
  • Create custom presets for repeated tasks

Why Choose Gemini 2.0?

The new Gemini models offer several compelling advantages:

  1. Unprecedented Cost-Effectiveness: Gemini 2.0 models are dramatically more affordable than competitors, offering up to 90% cost savings:

    • Processing 1M tokens with GPT-4 costs around $10
    • The same workload with Gemini 2.0 Flash costs just $0.40
    • Flash-Lite reduces costs even further at $0.30 per 1M tokens

    This cost advantage becomes especially significant for large-scale processing. For example, when analyzing large documents like technical documentation or research papers, Gemini can process 6,000 pages of PDFs with better accuracy than competitors at a fraction of the cost.

  2. Large Context Window: With up to 2M tokens context window in Pro, you can process massive documents or codebases in one go. That's equivalent to about 100,000 lines of code or 16 novels! This is significantly larger than competitors like GPT-4 and Claude, which are limited to 128K tokens.

  3. Multimodal Capabilities: All models support multiple input types, including text, images, audio, and video, making them versatile for various use cases.

Getting Started

Ready to try Gemini 2.0 models on your Mac? Download Kerlig↗ and experience the power of these new models with our native macOS interface. Our app makes it easy to leverage these advanced AI capabilities in your daily workflow, whether you're coding, writing, or analyzing content.

For more detailed information about using Gemini models with Kerlig, check out our help documentation↗.

References