Model Catalog & Sizing Guide

What Exactly is an "AI Model"?

An AI model is not a traditional computer program with lines of code telling it "if this, do that." Instead, it is a massive statistical mathematical network trained to understand relationships between words, concepts, and logic.

1. The Architecture

The structural blueprint. Modern Large Language Models (LLMs) use the Transformer architecture, which allows the AI to pay attention to multiple words across a page simultaneously rather than reading one word at a time.

2. The Weights (Parameters)

The learned knowledge. During training on billions of books and articles, the model calculates billions of numerical "weights" that store grammar, factual connections, reasoning steps, and idioms.

3. The Tokenizer

The translator. Computers cannot read letters; the tokenizer converts your human words into numbers (tokens), feeds them through the weights, and translates the output numbers back into fluent English.

What Are "Frontier Models"?

In the AI research world, the term "Frontier Model" refers to the largest, most capable systems operating at the absolute cutting edge of human technology. They represent state-of-the-art reasoning, vision, and problem-solving.

🌌 Frontier Model Hall of Fame

• OpenAI GPT-4o & o1: Massive multimodal reasoning models powering ChatGPT Plus.
• Anthropic Claude 3.5 Sonnet: Industry-leading coding, synthesis, and nuanced prose.
• Google Gemini 1.5 Pro: Million-token context window processing hours of video and whole codebases.
• Meta Llama 3.1 405B: The world's largest open-weights model, trained on 16,000 GPUs.

⚡ Why Frontier Models Live in the Cloud

A frontier model like Llama 3.1 405B has 405 billion parameters and requires over 800 Gigabytes of pure VRAM just to load into memory!

To run one query, it requires a room of 8 to 16 high-end enterprise server GPUs (such as NVIDIA H100s or B200s costing \$40,000 each), thousands of watts of power, and liquid cooling. No laptop in the world can run them.

Cloud vs. Local Models: The Honest Comparison

Should you use cloud chatbots (ChatGPT, Claude) or local models (via LocalPilot)? Both have unique superpowers. Here is how they compare across critical dimensions:

Feature	Cloud AI (ChatGPT / Claude)	Local AI (LocalPilot)
Where Your Data Goes	Sent over the internet to remote corporate data centers	Processed locally on your device via 127.0.0.1; does not leave your machine
Cost & Billing	\$20/month per user or ongoing pay-per-token API bills	Free and open source (GPLv3) • Zero subscriptions, zero API keys for open models
Offline Functionality	Requires active Wi-Fi or cellular connectivity	Works offline after initial setup and model download
Confidential / Sensitive Data	May be logged or reviewed according to provider policies	Local execution keeps draft text, private notes, and code on your device
Speed & Latency	Fast, but throttled during global peak traffic spikes	Immediate start, zero server queues, consistent speeds
World Knowledge Breadth	Encyclopedic recall across hundreds of billions of facts	Excellent for writing, synthesis, and reasoning; narrower trivia
Control & Ownership	Company can change, censor, or shut down models anytime	You own the file forever • No one can remove or restrict it

What Does "3B", "7B", or "8B" Mean?

The letter "B" stands for Billion Parameters. In simple terms, parameters are the connections in the AI's virtual neural network:

⚡ Compact Models (1B – 3B Parameters)

Lightning fast responses, uses only ~2GB to 3GB of RAM. Ideal for quick drafts, rewriting emails, meeting summaries, and everyday laptops.

Runs smoothly on almost any laptop made after 2018

🧠 Standard Models (7B – 9B Parameters)

The sweet spot for complex reasoning, logical deduction, long-form creative writing, and nuanced debate. Requires 8GB to 16GB of total system RAM.

Best for M1/M2/M3/M4 Macs or 16GB PCs

🚀 Frontier & High-End Models (14B – 123B Parameters) [v0.2.5]

For workstations and high-memory rigs with 32GB+ RAM or dedicated GPUs (Llama 3.3 70B, Qwen 2.5 72B, DeepSeek R1 70B, Mistral Large 123B). Matches frontier proprietary APIs completely offline.

Toggle with 'RAM: All Models' in TUI • Perfect for high-spec workstations

🎯

Pick by Your Goal (Which One Do You Need?)

⚡

Everyday Assistant

Proofreading, email drafts, instant Q&A on any computer.

→ Llama 3.2 (3B)

🔬

Deep Logic & Math

Trained on textbook quality data for academic reasoning.

→ Phi-3 Mini (3.8B)

🌐

Multilingual & Code

Unmatched accuracy in 29+ languages and code snippets.

→ Qwen 2.5 (3B)

✍️

Creative Writing

Long stories, complex essays, and rich nuance.

→ Llama 3.1 (8B)

Tier 1

Everyday Laptops (4 GB – 8 GB RAM)

✓ Official Google ~1.6 GB RAM

Gemma 2 (2B)

Google DeepMind\'s lightweight masterpiece. Blazing fast inference, exceptional reasoning, and low memory footprint.

Resource: ai.google.dev/gemma

Download Size: ~1.6 GB

Best for: Everyday chat, notes, quick answers

localpilot chat gemma2:2b

✓ Official DeepSeek ~1.1 GB RAM

DeepSeek R1 Distill (1.5B)

DeepSeek AI\'s open reasoning distilled model. Features chain-of-thought mathematical reasoning in an ultra-compact package.

Resource: DeepSeek-R1

Download Size: ~1.1 GB

Best for: Math, logic, step-by-step reasoning

localpilot chat deepseek-r1:1.5b

✓ Official Qwen ~1.0 GB RAM

Qwen 2.5 Coder (1.5B)

Alibaba\'s specialized coding model. Code completion, debugging, script generation, and terminal explanations.

Resource: qwenlm.github.io

Download Size: ~1.0 GB

Best for: Python, Bash, web code, bug fixing

localpilot chat qwen2.5-coder:1.5b

Meta • Q4_K_M ~2.0 GB RAM

Llama 3.2 (3B)

Meta's flagship compact model. Incredibly fast, great at summarization, proofreading, and natural dialogue.

Download Size: ~2.0 GB

Download Time: ~2-3 min (100Mbps)

Best for: Emails, notes, daily Q&A

localpilot chat llama3.2:3b

Alibaba • Q4_K_M ~2.1 GB RAM

Qwen 2.5 (3B)

Extremely smart instruction-following and strong multilingual capabilities. Perfect for non-English inquiries and programming.

Download Size: ~2.1 GB

Download Time: ~2-3 min (100Mbps)

Best for: Multilingual & coding

localpilot chat qwen2.5:3b

Microsoft • Q4_K_M ~2.3 GB RAM

Phi-3 Mini (3.8B)

Microsoft's reasoning powerhouse. Trained on curated textbook-quality data for logic and academic questions.

Download Size: ~2.3 GB

Download Time: ~2-4 min (100Mbps)

Best for: Math, logic, science Q&A

localpilot chat phi3:mini

Tier 2

Balanced Workstations (8 GB – 16 GB RAM)

✓ Official DeepSeek ~4.7 GB RAM

DeepSeek R1 Distill (7B)

DeepSeek AI\'s full 7B distilled reasoning model. Deep thinking, chain-of-thought mathematical solutions, and competitive logic.

Resource: DeepSeek-R1

Download Size: ~4.7 GB

Best for: Complex math, logic, reasoning

localpilot chat deepseek-r1:7b

✓ Official Qwen ~4.7 GB RAM

Qwen 2.5 Coder (7B)

Flagship 7B coding model. Code architecture, unit test writing, complex refactoring, and multi-file debugging.

Resource: qwenlm.github.io

Download Size: ~4.7 GB

Best for: Software development, algorithms

localpilot chat qwen2.5-coder:7b

✓ Official Zhipu AI ~5.5 GB RAM

GLM-4 (9B)

Zhipu AI & THUDM open bilingual powerhouse. Exceptionally strong in both Chinese and English comprehension and synthesis.

Resource: THUDM/GLM-4

Download Size: ~5.5 GB

Best for: Bilingual tasks, reasoning, translation

localpilot chat glm4:9b

Meta • Q4_K_M ~4.7 GB RAM

Llama 3.1 (8B)

The gold standard for general intelligence. Excellent storytelling, research analysis, and creative ideation.

Download Size: ~4.7 GB

Download Time: ~4-7 min (100Mbps)

Best for: Creative writing & research

localpilot chat llama3.1:8b

Mistral AI • Q4_K_M ~4.1 GB RAM

Mistral (7B)

Renowned European model. Crisp, concise responses, excellent technical explanations and structured formatting.

Download Size: ~4.1 GB

Download Time: ~3-6 min (100Mbps)

Best for: Concise summaries & code

localpilot chat mistral:7b

Google • Q4_K_M ~5.4 GB RAM

Gemma 2 (9B)

Google's high-efficiency open model. Outstanding benchmarks for factual knowledge, reading comprehension, and math.

Download Size: ~5.4 GB

Download Time: ~5-9 min (100Mbps)

Best for: Fact-checking & academics

localpilot chat gemma2:9b

🔒

Model Safety & Provenance (Where Do They Come From?)

When you download a model with LocalPilot, where does it actually come from, and can it contain viruses?

1. Official Sources

All model files are retrieved directly from verified public repositories: the official Ollama registry or Hugging Face GGUF repositories created by recognized labs (Meta, Google, Microsoft).

2. SHA-256 Checksums

LocalPilot verifies the cryptographic fingerprint (SHA-256 hash) of the downloaded file whenever available. If even a single byte is corrupted or altered during download, it rejects the file immediately.

3. Static Data Files

A model file (like a .gguf) contains math matrices (numbers), not executable program code. It cannot run arbitrary software on your computer or alter your personal files.

🛡️

The LocalPilot Safety Shield

You never have to calculate memory math manually. Whenever you run localpilot discover, our hardware scanner inspects your available RAM and marks models with a safe compatibility percentage so your system never runs out of memory.

Everything About AI Models: Frontier, Cloud & Local

What Exactly is an "AI Model"?

What Are "Frontier Models"?

Cloud vs. Local Models: The Honest Comparison

What Does "3B", "7B", or "8B" Mean?

Pick by Your Goal (Which One Do You Need?)

Everyday Assistant

Deep Logic & Math

Multilingual & Code

Creative Writing

Everyday Laptops (4 GB – 8 GB RAM)

Gemma 2 (2B)

DeepSeek R1 Distill (1.5B)

Qwen 2.5 Coder (1.5B)

Llama 3.2 (3B)

Qwen 2.5 (3B)

Phi-3 Mini (3.8B)

Balanced Workstations (8 GB – 16 GB RAM)

DeepSeek R1 Distill (7B)

Qwen 2.5 Coder (7B)

GLM-4 (9B)

Llama 3.1 (8B)

Mistral (7B)

Gemma 2 (9B)

Model Safety & Provenance (Where Do They Come From?)

The LocalPilot Safety Shield