Local Models for AI Coding

Цей контент ще не доступний вашою мовою.

AI/ML Engineering Track | Complexity: [MEDIUM] | Time: 3-4

Or: How to Stop Paying OpenAI and Start Loving Your GPU

Reading Time: 3-4 hours Prerequisites: Module 1.1 complete, 16GB+ RAM recommended (8GB minimum)

The $600/Year Mistake That Changed Everything

Austin, Texas. March 15, 2024. 8:43 PM.

Marcus stared at his credit card statement in disbelief. Three hundred and twelve dollars. In a single month. All from OpenAI.

He’d been working on a side project—a code analysis tool—and had been iterating rapidly with gpt-5. Every experiment, every prompt tweak, every debugging session had added up. The tokens had accumulated silently while he was in the flow of coding.

“There has to be another way,” he muttered, opening a new browser tab.

That night, Marcus discovered Ollama. By midnight, he had DeepSeek Coder running on his MacBook. The responses were slower, sure, but the quality was surprisingly good. And the cost? Zero. He ran the same prompts that had cost him fifty dollars the day before—completely free.

Two months later, Marcus’s approach had evolved. He used local models for 80% of his work: code completion, test generation, documentation, routine refactoring. For the hard problems—complex architecture decisions, tricky bugs—he’d spend a few dollars on Claude. His monthly API bill dropped from $300 to $15.

“Local models aren’t a replacement for gpt-5 or Claude. They’re more like a capable junior developer who handles the routine work so your expensive senior developer can focus on what matters. Once you think about it that way, the hybrid approach becomes obvious.” — Marcus Chen, at PyCon 2025

That $600/year in savings? It funded Marcus’s new GPU, which made his local models even faster. The cycle of value continued.

What You’ll Be Able to Do

By the end of this module, you will:

Understand the local vs API trade-offs for AI coding
Install and run Ollama with coding-focused models
Use local models with Aider for terminal-based AI coding
Configure Continue.dev to use local models in VS Code
Implement cost-optimization strategies (local + API hybrid)
Know which local models excel at which coding tasks
Troubleshoot common local model issues

Why this matters: Running models locally means $0/month AI coding, complete privacy, and offline capability. Perfect for learning, experimentation, and cost-conscious development.

The Local Models Revolution: Why This Changes Everything

The Problem Everyone Faces

Without local models (API-only approach):

Week 1: Sign up for Claude API
Week 2: $10 in API costs (moderate coding)
Week 3: $15 (got ambitious with refactoring)
Week 4: $25 (trying different models, debugging)

Monthly total: $50-100 in API costs
Annual: $600-1,200

Plus:
- Need internet connection always
- Code sent to external servers
- Rate limits during heavy use
- Billing surprises

With local models (hybrid approach):

Week 1: Install Ollama (free), download DeepSeek Coder (free)
Week 2: Use local for 80% of tasks, API for complex 20%
Week 3: API costs: $3 (only used for hard problems)
Week 4: API costs: $2

Monthly total: $5-10 in API costs
Annual: $60-120 (10× cheaper!)

Plus:
- Works offline
- Code stays on your machine
- No rate limits
- Predictable costs

The difference: $600-1,200/year vs $60-120/year. 10× cost reduction!

Did You Know? The Open Source AI Boom

2023: OpenAI dominates, local models struggle to compete

2024: Meta releases Llama 4 - changes everything

Open source models catch up dramatically
DeepSeek Coder beats GPT-3.5 on coding tasks
Qwen 2.5-Coder rivals commercial models

2025: Local models are now viable alternatives

DeepSeek R1 competes with o1 on reasoning
Qwen 2.5-Coder:32B rivals gpt-5 on code
Can run 7B models on MacBook Air!

The trend: Gap between local and API models shrinking fast. By 2026, local models may match gpt-5 quality.

Introduction: Understanding Local Models

What Are Local Models?

The Personality: Local models are self-sufficient coding assistants - they live on your machine, not in the cloud.

Think of local models like having a reference library in your home versus using the public library downtown. The public library (API models) has more books, expert librarians, and the latest publications—but you have to drive there, pay parking, and work within their hours. Your home library (local models) has fewer books, but it’s always available, completely private, and costs nothing after the initial purchase.

Three types of AI model deployment:

1. API Models (ChatGPT, Claude, Gemini)
   ├─ Run on company servers
   ├─ Access via internet
   ├─ Pay per request
   └─ Example: claude-4.6-sonnet-20241022

2. Local Models (Qwen, DeepSeek, Llama)
   ├─ Downloaded to your machine
   ├─ Run on your CPU/GPU
   ├─ One-time download cost
   └─ Example: qwen2.5-coder:7b

3. Hybrid (Mix of both)
   ├─ Local for simple/repetitive tasks
   ├─ API for complex reasoning
   └─ Best of both worlds

Scale analogy:

API models = Cloud computing (AWS, Google Cloud)
Local models = Your own server
Hybrid = Edge computing + cloud

Process analogy:

API models = Calling a consultant (pay per hour, expert knowledge)
Local models = Hiring an intern (one-time training, available 24/7)
Hybrid = Small team with occasional expert advice

Local vs API: When to Use Each

Use Local Models When:

1. Cost is a concern ($0/month vs $50-100/month)

# Scenario: Learning Python, writing 100+ functions/week
# API cost: $20-40/month (Claude/gpt-5)
# Local cost: $0/month (after download)
# Annual savings: $240-480

2. Privacy matters (proprietary code)

# Scenario: Working on closed-source company project
# API: Code sent to Anthropic/OpenAI servers
# Local: Code never leaves your machine

3. Offline work needed

# Scenario: Coding on airplane, no WiFi
# API: Cannot work
# Local: Full functionality

4. Repetitive tasks (boilerplate, tests, docs)

# Scenario: Generating 50 unit tests
# API: $2-5 in costs
# Local: Free, unlimited

5. Learning and experimentation

# Scenario: Trying different prompts, models
# API: Costs add up with experimentation
# Local: Experiment freely

Use API Models When:

1. Maximum quality needed

# Scenario: Complex algorithm design, architecture decisions
# API: Claude Opus 4, gpt-5 (best reasoning)
# Local: Good, but not as strong

2. Latest capabilities required

# Scenario: Using newest features (vision, long context, etc.)
# API: Always cutting-edge
# Local: Lags behind by 6-12 months

3. Very large context windows

# Scenario: Analyzing entire 50K line codebase
# API: Claude (200K tokens), Gemini (2M tokens)
# Local: Most models max at 8-32K tokens

4. Zero setup time

# Scenario: Need to start coding NOW
# API: Sign up, get key, start (5 minutes)
# Local: Download (30min-2hrs depending on model size)

The Decision Matrix

Factor	Local Models	API Models	Winner
Cost	$0/month	$20-100/month	Local
Privacy	Code stays local	Sent to servers	Local
Offline	Works offline	Needs internet	Local
Quality	Good	Excellent	API
Speed	Slower (CPU)	Very fast	API
Context	8-32K tokens	200K-2M tokens	API
Latest features	6-12mo lag	Cutting-edge	API
Setup	30min-2hrs	5 minutes	API

Verdict: Use both! Local for most work, API for complex tasks.

Think of the hybrid approach like a restaurant kitchen. You don’t fly in a Michelin-starred chef (expensive API model) to chop onions and wash dishes—that’s what your reliable prep cooks (local models) are for. But when it’s time to create the signature dish that brings customers back, you want the master chef’s expertise. The kitchen runs best when everyone works together, each handling what they do best.

The Local Model Landscape (2025)

Category 1: Best for Coding Quality ⭐⭐⭐

DeepSeek Coder V2 (China - DeepSeek AI)

What it is: Specialized coding model that rivals gpt-5 on benchmarks

The Good:

Best code quality among local models
Supports 338 programming languages
Strong at debugging and refactoring
Reasoning capabilities (R1 model)
16B version runs on 16GB RAM

The Not-So-Good:

️ Newer, less community support
️ 236B version needs GPU
️ Less documentation than Llama

Sizes available:

ollama pull deepseek-coder-v2:16b   # Recommended
ollama pull deepseek-coder-v2:236b  # Needs GPU
ollama pull deepseek-r1:7b          # Reasoning model
ollama pull deepseek-r1:14b

Best for: Complex algorithms, code review, refactoring Benchmarks: Beats GPT-3.5-Turbo, rivals gpt-5 on HumanEval

** Did You Know? The DeepSeek Story**

DeepSeek is a Chinese AI company that shocked Silicon Valley in 2024-2025. While American AI labs spent billions on massive GPU clusters, DeepSeek’s small team in Hangzhou achieved comparable results with a fraction of the compute budget.

Their secret? Engineering efficiency over brute force. DeepSeek R1 (their reasoning model) was trained with novel techniques that reduced compute costs by 90% compared to OpenAI’s o1.

The result: A model you can run locally that competes with $200/month API subscriptions. DeepSeek proved that AI progress isn’t just about throwing money at problems - clever engineering matters more.

Fun fact: When DeepSeek R1 was released in January 2025, Nvidia’s stock dropped 17% ($600B in market cap) in one day. Investors realized the AI hardware arms race might be less important than they thought.

Qwen 2.5-Coder (China - Alibaba)

What it is: Latest coding model from Alibaba, excellent quality

The Good:

Top-tier code generation
Fast inference
Great at multiple languages (especially Asian languages)
Strong reasoning (QwQ model)
7B version very efficient

The Not-So-Good:

️ Less known in Western dev community
️ Some English docs still being translated

Sizes available:

ollama pull qwen2.5-coder:7b    # Best balance
ollama pull qwen2.5-coder:14b
ollama pull qwen2.5-coder:32b   # High quality
ollama pull qwq:32b             # Reasoning

Best for: Code generation, multi-language projects, daily coding Benchmarks: Often beats CodeLlama, competitive with GPT-3.5

Category 2: Most Popular & Mature ⭐

CodeLlama (USA - Meta)

What it is: Meta’s specialized coding version of Llama 4

The Good:

Huge community, best documentation
Well-tested and stable
Great Python support
Multiple size options
Excellent code completion

The Not-So-Good:

️ Being overtaken by newer models (DeepSeek, Qwen)
️ Not as strong at reasoning

Sizes available:

ollama pull codellama:7b
ollama pull codellama:13b
ollama pull codellama:34b

Best for: Beginners, Python development, stable choice Benchmarks: Solid, but DeepSeek and Qwen now beat it

Category 3: Fastest Inference

Codestral (France - Mistral AI)

What it is: Mistral’s coding-specialized model

The Good:

Very fast inference
22B params, good quality
Excellent for autocomplete
Strong European backing

The Not-So-Good:

️ 22B size needs more RAM
️ Not as widely used as Llama

Sizes available:

ollama pull codestral:22b

Best for: Real-time autocomplete, fast iteration Benchmarks: Competitive with CodeLlama 34B

Category 4: Smallest & Most Efficient

Phi-3.5 (USA - Microsoft)

What it is: Microsoft’s tiny but capable model

The Good:

Only 3.8B params - runs anywhere!
Surprisingly good for size
Great for learning
Fast on CPU

The Not-So-Good:

️ Limited capabilities vs larger models
️ Small context window
️ Not suitable for complex tasks

Sizes available:

ollama pull phi3:3.8b
ollama pull phi3.5:3.8b

Best for: Low-resource machines, learning, quick tasks Benchmarks: Punches above its weight class

Category 5: Google’s Offering

CodeGemma (USA - Google)

What it is: Google’s open-source coding model

The Good:

Google quality
Good Python support
Active development

The Not-So-Good:

️ Less popular than Llama/Qwen
️ Smaller community

Sizes available:

ollama pull codegemma:7b
ollama pull gemma2:9b
ollama pull gemma2:27b

Best for: Google ecosystem, Python-focused work

Comparison Table: Which Model Should You Use?

Model	Size	RAM Needed	Code Quality	Speed	Best Use Case
DeepSeek Coder V2	16B	16GB	⭐⭐⭐⭐⭐	Medium	Complex coding, refactoring
Qwen 2.5-Coder	7B	8GB	⭐⭐⭐⭐⭐	Fast	Daily driver, general coding
Qwen 2.5-Coder	32B	32GB	⭐⭐⭐⭐⭐	Medium	High-quality code gen
CodeLlama	13B	16GB	⭐⭐⭐⭐	Fast	Python, stable choice
Codestral	22B	24GB	⭐⭐⭐⭐	Very Fast	Autocomplete, real-time
CodeGemma	7B	8GB	⭐⭐⭐	Fast	Python-focused
Phi-3.5	3.8B	4GB	⭐⭐⭐	Very Fast	Learning, low-resource

Recommendations by Your Situation

You Have: MacBook with 16GB RAM (Most Common)

# Your best setup:
ollama pull qwen2.5-coder:7b          # Daily driver (uses ~6GB)
ollama pull deepseek-coder-v2:16b     # Heavy lifting (uses ~14GB)

Why: Qwen 7B for fast tasks, DeepSeek 16B when you need quality

You Have: MacBook with 8GB RAM

# Your best setup:
ollama pull qwen2.5-coder:7b          # Main model
ollama pull phi3.5:3.8b               # Quick tasks

Why: Stay under 8GB, both are excellent for their size

You Have: Desktop with 32GB+ RAM (or GPU)

# Go big!
ollama pull qwen2.5-coder:32b         # Highest quality
ollama pull deepseek-coder-v2:236b    # If you have GPU

Why: Use the full power available

You Have: Budget Laptop (4GB RAM)

# Lightweight only:
ollama pull phi3.5:3.8b

Why: Only model that will run smoothly on 4GB

Did You Know? The Apple Silicon Revolution

Why Your MacBook Can Run AI Now

Before 2020, running AI models locally was painful. You needed expensive Nvidia GPUs, separate RAM for graphics, and complex CUDA drivers. Most developers couldn’t do it.

Then Apple released the M1 chip in November 2020. Everything changed.

The breakthrough: Apple’s “Unified Memory Architecture” means the CPU and GPU share the same fast memory. For AI inference, this is perfect - no copying data back and forth between CPU and GPU RAM.

The numbers:

M1 MacBook Air (8GB): Can run 7B parameter models
M1 Pro (16GB): Can run 14B parameter models comfortably
M1 Max (32GB): Can run 32B+ parameter models
M1 Ultra (64-128GB): Can run models that normally need datacenter GPUs

The surprise: A $999 MacBook Air can now run models that would have cost $10,000+ in GPU hardware just 3 years ago.

Why this matters for you: If you have any Mac with M-series chip (M1, M2, M3, M4), you can run production-quality AI models locally. No cloud costs, complete privacy, works offline. This was science fiction in 2019.

️ Hands-On: Installing Ollama

What is Ollama?

The Personality: Ollama is your model manager - like Docker for AI models.

Think of Ollama like a video game console for AI models. Just like how a PlayStation lets you download, manage, and play games without worrying about hardware compatibility or installation headaches, Ollama handles all the complexity of running AI models. You just say “I want this model,” and Ollama figures out the memory management, optimization, and APIs automatically.

What it does:

Downloads and manages models
Runs models locally
Provides simple API
Works with coding tools (Aider, Continue.dev)

** Did You Know?**

Ollama was created by Jeffrey Morgan and Michael Chiang in 2023. They were frustrated that running local AI models required deep knowledge of CUDA drivers, quantization formats, and memory optimization. Their goal was simple: make local models as easy as docker run. Within a year, Ollama had over 500,000 users and became the de facto standard for running local models. The name “Ollama” is a playful take on “llama” (Meta’s model family) combined with the idea of models being “portable” (O-llama, like a friendly llama you can carry around).

Installation (macOS)

# Method 1: Homebrew (recommended)
brew install ollama

# Method 2: Official installer
# Download from https://ollama.com/download

# Verify installation
ollama --version
# Output: ollama version 0.x.x

Installation (Linux)

# Single command install
curl -fsSL https://ollama.com/install.sh | sh

# Verify
ollama --version

# Start service (if needed)
sudo systemctl start ollama
sudo systemctl enable ollama

Installation (Windows)

# Download installer from https://ollama.com/download
# Run OllamaSetup.exe

# Verify in PowerShell
ollama --version

Your First Model

# Pull a small model to test
ollama pull qwen2.5-coder:7b

# This downloads ~4GB, takes 5-15 minutes depending on connection
# Output:
# pulling manifest
# pulling 8934d96d3f08... 100% ▕████████████▏ 4.7 GB
# pulling 8c17c2ebb0ea... 100% ▕████████████▏ 7.0 KB
# pulling 590d74a5569b... 100% ▕████████████▏ 6.0 KB
# pulling 56bb8bd477a5... 100% ▕████████████▏  96 B
# pulling 1d21db061cdd... 100% ▕████████████▏ 485 B
# success

Test Your Model

# Interactive mode
ollama run qwen2.5-coder:7b

# You'll see a prompt:
# >>>

# Try it:
>>> Write a Python function to reverse a string

# Model generates code!
# Exit with: /bye

List Your Models

# See what models you have
ollama list

# Output:
# NAME                    ID              SIZE      MODIFIED
# qwen2.5-coder:7b        abc123          4.7 GB    2 minutes ago

Remove Models (Free Up Space)

# Remove a model
ollama rm codellama:13b

# Check space saved
ollama list

Hands-On: Using Local Models with Aider

What is Aider?

Aider = AI pair programmer in your terminal

Edits multiple files
Git integration
Works with any LLM (API or local!)

Installing Aider

# Install via pip
pip install aider-chat

# Or in your venv
source venv/bin/activate
pip install aider-chat

# Verify
aider --version

Using Aider with Local Models

# Basic usage with Qwen
aider --model ollama/qwen2.5-coder:7b

# With specific files
aider --model ollama/qwen2.5-coder:7b myfile.py

# With DeepSeek for complex work
aider --model ollama/deepseek-coder-v2:16b

Example Session

# Start Aider with local model
aider --model ollama/qwen2.5-coder:7b test.py

# Aider opens, you see:
# Aider v0.x.x
# Model: ollama/qwen2.5-coder:7b
# Git repo: /path/to/project
# Added test.py to the chat

# Now ask it to code:
> Create a Python function to calculate fibonacci numbers with memoization

# Aider generates code:
# Applied edit to test.py
# Commit changes? (Y)es/(N)o/(D)on't ask again [Yes]: y
# Commit message: Add fibonacci function with memoization

# Success! Code written, tests added, committed to git

Aider Configuration File

Create ~/.aider.conf.yml:

# Use local model by default
model: ollama/qwen2.5-coder:7b

# Auto-commit changes
auto-commits: true

# Show diffs
show-diffs: true

# Prettier output
pretty: true

Now just run:

aider  # Uses config automatically

Cost Comparison: Aider Local vs API

Scenario: Refactor 10 files (500 lines each)

With Claude API:

Input: ~100K tokens
Output: ~50K tokens
Cost: ~$5-8 per refactoring session
Monthly (4 sessions): $20-32

With Local Qwen:

Input: Unlimited
Output: Unlimited
Cost: $0
Monthly (unlimited sessions): $0

Savings: $20-32/month = $240-384/year!

Hands-On: Using Local Models with Continue.dev

What is Continue.dev?

Continue.dev = VS Code extension for AI coding

Like Copilot, but works with ANY model
Supports local models!
Open source

Installing Continue.dev

Open VS Code
Extensions (Cmd/Ctrl+Shift+X)
Search “Continue”
Install “Continue - Codestral, Claude, and more”

Configuring for Local Models

Open Continue settings (Cmd/Ctrl+Shift+J)
Click gear icon ️
Edit config.json:

{
  "models": [
    {
      "title": "Qwen 2.5 Coder (Local)",
      "provider": "ollama",
      "model": "qwen2.5-coder:7b",
      "apiBase": "http://localhost:11434"
    },
    {
      "title": "DeepSeek Coder (Local)",
      "provider": "ollama",
      "model": "deepseek-coder-v2:16b",
      "apiBase": "http://localhost:11434"
    }
  ],
  "tabAutocompleteModel": {
    "title": "Qwen Autocomplete",
    "provider": "ollama",
    "model": "qwen2.5-coder:7b"
  }
}

Save and reload VS Code

Using Continue with Local Models

Tab Autocomplete:

Start typing code
Continue suggests completions
Press Tab to accept
All running locally!

Chat:

Open Continue sidebar (Cmd/Ctrl+Shift+J)
Select “Qwen 2.5 Coder (Local)”

Ask questions:

You: Refactor this function to use async/await

Qwen: [Generates refactored code]

Inline Editing:

Highlight code
Cmd/Ctrl+I
Describe change: “Add error handling”
Continue edits inline!

Multi-Model Setup (Best Practice)

Use different models for different tasks:

{
  "models": [
    {
      "title": "Quick (Phi-3.5)",
      "provider": "ollama",
      "model": "phi3.5:3.8b",
      "description": "Fast for simple tasks"
    },
    {
      "title": "Balanced (Qwen)",
      "provider": "ollama",
      "model": "qwen2.5-coder:7b",
      "description": "Daily driver"
    },
    {
      "title": "Quality (DeepSeek)",
      "provider": "ollama",
      "model": "deepseek-coder-v2:16b",
      "description": "Complex refactoring"
    },
    {
      "title": "API (Claude)",
      "provider": "anthropic",
      "model": "claude-4.6-sonnet-20241022",
      "apiKey": "YOUR_KEY",
      "description": "When you need the best"
    }
  ]
}

Switch models based on task complexity!

Cost Optimization: Hybrid Strategy

The 80/20 Rule

80% of tasks: Simple, use local models (FREE)

Code completion
Boilerplate generation
Unit tests
Documentation
Simple refactoring

20% of tasks: Complex, use API (PAY)

Architecture design
Complex algorithms
Debugging subtle bugs
Code review of large changes

Result: 80% cost reduction!

Your Hybrid Setup (Recommended)

For daily coding:

# VS Code + Continue.dev
# Autocomplete: Qwen 2.5-Coder:7b (local)
# Chat: Mix of local (simple) + Gemini Flash (complex, your free tier)

For terminal work:

# Aider with local
aider --model ollama/qwen2.5-coder:7b

# When stuck, switch to API
aider --model gemini/gemini-2.5-flash

Monthly cost:

Local models: $0
Gemini Flash (free tier): $0
Overflow to paid API: $3-5/month

Total: ~$5/month vs $50-100 API-only!

Performance Benchmarks

Code Generation Quality

HumanEval Benchmark (Python coding tasks):

Model	Score	Cost/1M tokens	Speed
gpt-5	67%	$30	Fast
Claude 3.5 Sonnet	64%	$15	Fast
DeepSeek Coder V2 16B	62%	$0	Medium
Qwen 2.5-Coder 32B	61%	$0	Medium
CodeLlama 34B	54%	$0	Medium
Gemini Flash 2.5	71%	$0 (free tier)	Very Fast

Insight: Local models are 90-95% as good as API, at 0% of the cost!

Speed Comparison (MacBook Pro 16GB)

Task: Generate 100-line Python function

Model	Time	Quality
gpt-5 (API)	2-3 sec	⭐⭐⭐⭐⭐
Qwen 2.5-Coder:7b (local)	8-12 sec	⭐⭐⭐⭐
DeepSeek V2:16b (local)	15-20 sec	⭐⭐⭐⭐⭐
Phi-3.5:3.8b (local)	5-8 sec	⭐⭐⭐

Trade-off: Local is 3-6× slower, but free and private!

Common Mistakes: Learn From Others’ Pain

Mistake #1: “Downloaded 236B Model, My Mac is Frozen”

Symptom: Downloaded DeepSeek 236B on 16GB MacBook, system unresponsive

Why It’s Bad:

236B model needs 200+ GB RAM or GPU
CPU-only inference would take minutes per response
System swaps to disk, grinds to halt

The Fix:

# Remove the huge model
ollama rm deepseek-coder-v2:236b

# Use appropriately-sized model
ollama pull deepseek-coder-v2:16b  # Much better!

Rule: Match model size to your RAM (model params × 2 = GB needed)

Mistake #2: “Local Model Gives Worse Code Than I Expected”

Symptom: Qwen 7B generates buggy code, frustrated

Why It’s Bad:

Used 7B model for complex architecture task
7B models are good, but not great at complex reasoning

The Fix:

# For complex tasks, use bigger model OR API
aider --model ollama/deepseek-coder-v2:16b  # Better
# or
aider --model gemini/gemini-2.5-flash       # Your free tier

Rule: Match model capability to task complexity!

Mistake #3: “Ollama Using 100% CPU, Laptop is Hot”

Symptom: Fan running loud, laptop hot during code generation

Why It’s Normal:

Local models use CPU/GPU intensively
This is expected behavior
Not harmful unless sustained for hours

The Fix (if bothered):

# Use smaller model for less intensive work
ollama pull phi3.5:3.8b

# Or use API for complex stuff
# Save local for offline/simple tasks

Rule: Local models trade electricity for cost savings!

Mistake #4: “Continue.dev Not Finding My Local Model”

Symptom: Set up Ollama, but Continue.dev says “model not found”

Diagnosis:

# Check Ollama is running
ollama list

# If empty or error, Ollama service isn't running

The Fix:

# Start Ollama service (macOS/Linux)
ollama serve

# Or on Linux:
sudo systemctl start ollama

# Then in Continue config:
# Make sure apiBase is: http://localhost:11434

Prevention: Always verify ollama list works before configuring tools!

Best Practices for Local Models

1. Start Small, Scale Up

Why: Don’t download every model at once

# BAD (downloading everything)
ollama pull qwen2.5-coder:7b
ollama pull qwen2.5-coder:14b
ollama pull qwen2.5-coder:32b
ollama pull deepseek-coder-v2:16b
ollama pull codellama:34b
# (Uses 80+ GB disk space!)

# GOOD (start with one)
ollama pull qwen2.5-coder:7b
# Test it for a week
# If you need more quality, then add:
ollama pull deepseek-coder-v2:16b

Benefit: Save disk space, learn what works for you

2. Use Local for Iteration, API for Innovation

Pattern:

# Scenario: Building new feature

# Step 1: Architecture (use API - needs reasoning)
aider --model gemini/gemini-2.5-flash
> "Design class structure for user authentication system"

# Step 2: Implementation (use local - repetitive)
aider --model ollama/qwen2.5-coder:7b
> "Implement the User class with methods we designed"

# Step 3: Tests (use local - boilerplate)
> "Generate unit tests for all User methods"

# Step 4: Refactoring (use API if complex, local if simple)
aider --model ollama/deepseek-coder-v2:16b
> "Refactor for better error handling"

Benefit: Best quality where it matters, zero cost where it doesn’t

3. Create Model Aliases for Workflows

Setup (in ~/.bashrc or ~/.zshrc):

# Quick coding tasks
alias aider-quick="aider --model ollama/phi3.5:3.8b"

# Daily driver
alias aider-local="aider --model ollama/qwen2.5-coder:7b"

# High quality
alias aider-quality="aider --model ollama/deepseek-coder-v2:16b"

# API fallback (your Gemini free tier)
alias aider-api="aider --model gemini/gemini-2.5-flash"

Usage:

# Simple task
aider-quick add_tests.py

# Normal coding
aider-local main.py

# Complex refactoring
aider-quality --architect

Benefit: Fast workflow switching!

4. Monitor Your Costs (Even $0 Has Opportunity Cost)

Track usage:

# Create usage log
echo "$(date): Used DeepSeek 16B for refactoring - 5 min CPU time" >> ~/ai_usage.log

# Weekly review
cat ~/ai_usage.log | grep "$(date +%Y-%m)"

Calculate savings:

# If you used API instead
api_cost_per_1m_tokens = 3  # Claude Haiku
estimated_tokens = 50_000   # Your usage
monthly_savings = (estimated_tokens / 1_000_000) * api_cost_per_1m_tokens * 30
print(f"Saved ${monthly_savings:.2f} this month by using local models")

Benefit: Quantify your cost optimization!

Deep Dive (Optional): How Local Models Work

For the curious: What happens when you run ollama run qwen2.5-coder:7b?

The Technical Stack

1. Model Download:

Ollama downloads:
├─ Model weights (4.7 GB) - the "brain"
├─ Tokenizer (7 KB) - converts text to numbers
├─ Config (6 KB) - model parameters
└─ Prompt template (485 B) - how to format inputs

2. Loading into RAM:

7B model × 2 bytes/param = 14 GB quantized
Actually uses: 4-6 GB (thanks to quantization!)

3. Inference:

Your prompt → Tokenizer → Model → Detokenizer → Code output

Example:
"Write Python function" → [1234, 5678, 9012] → Transformer layers → [4567, 8901, 2345] → "def fibonacci(n):"

4. Quantization Magic:

Full precision: 16-bit floats = 14 GB
Quantized (Q4): 4-bit ints = 3.5 GB
Quantized (Q8): 8-bit ints = 7 GB

Ollama uses Q4 by default - 4× memory reduction!

5. CPU vs GPU:

CPU inference: Uses all cores, ~10-20 tokens/sec
GPU inference (CUDA): Uses GPU, ~50-100 tokens/sec
Apple Silicon (Metal): Uses GPU, ~30-60 tokens/sec

Try This: Interactive Challenges

Challenge 1: The Model Comparison Test

Goal: See which model is best for YOUR coding style

Task: Generate the same function with 3 different models, compare

# Create test file
cat > compare_models.md << 'EOF'
# Model Comparison
Task: Write a Python function to find prime numbers up to N using Sieve of Eratosthenes

## Model 1: Qwen 2.5-Coder:7b
[Paste output here]

## Model 2: DeepSeek Coder V2:16b
[Paste output here]

## Model 3: CodeLlama:13b
[Paste output here]

## Winner: [Your choice]
Why: [Your reasoning]
EOF

# Test each model
ollama run qwen2.5-coder:7b "Write Python function for Sieve of Eratosthenes"
ollama run deepseek-coder-v2:16b "Write Python function for Sieve of Eratosthenes"
ollama run codellama:13b "Write Python function for Sieve of Eratosthenes"

Expected Results

Likely winner: DeepSeek V2:16b (most optimized, best comments) Runner-up: Qwen 2.5-Coder:7b (fast, correct, good enough) Lesson: Bigger models produce slightly better code, but 7B is often sufficient!

Challenge 2: Cost Savings Calculator

Goal: Calculate your actual savings from using local models

Task: Track one week of coding, calculate what it would cost with API

# Create savings_calculator.py
def calculate_savings():
    # Your actual usage this week
    coding_sessions = 15  # How many times you used Aider/Continue
    avg_tokens_per_session = 5000  # Rough estimate

    total_tokens = coding_sessions * avg_tokens_per_session

    # API costs
    claude_sonnet_cost = (total_tokens / 1_000_000) * 3  # $3/M input
    gemini_flash_cost = 0  # Your free tier

    # Local cost
    local_cost = 0

    weekly_savings = claude_sonnet_cost
    monthly_savings = weekly_savings * 4
    annual_savings = monthly_savings * 12

    print(f" Your Cost Savings Report")
    print(f"   Weekly:  ${weekly_savings:.2f}")
    print(f"   Monthly: ${monthly_savings:.2f}")
    print(f"   Annual:  ${annual_savings:.2f}")
    print(f"\n    You saved enough for: [insert fun comparison]")

calculate_savings()

Run it:

python savings_calculator.py

Expected Output

 Your Cost Savings Report
   Weekly:  $0.23
   Monthly: $0.90
   Annual:  $10.80

    You saved enough for: 1 month of Spotify Premium!

Note: Actual savings depend on usage. Heavy users save $200-500/year!

Module 1.2 Complete Checklist

Use this to verify you’re ready for Module 2:

Setup

Ollama installed and working (ollama --version)
At least one coding model downloaded (Qwen or DeepSeek)
Tested model in terminal (ollama run qwen2.5-coder:7b)
Aider installed (aider --version)
Continue.dev extension installed in VS Code

Practice

Generated code with local model (any task)
Used Aider with local model
Configured Continue.dev to use local model
Tried at least 2 different models, compared results
Set up hybrid workflow (local + Gemini Flash)

Understanding

Can explain local vs API trade-offs
Know which model to use for which task
Understand cost savings (calculated your own)
Know when to use local vs when to switch to API

Reflection

Identified 3 tasks you’ll do with local models
Calculated potential monthly savings
Planned your hybrid workflow (80% local, 20% API)

All checked? You’re ready for Module 2: Prompt Engineering!

️ Next Steps

Congratulations! You now have cost-effective AI coding with local models!

You’ve learned:

How to install and run Ollama
The best local models for coding (DeepSeek, Qwen, etc.)
Using local models with Aider
Configuring Continue.dev for local models
Hybrid optimization strategies (local + API)
Real cost savings ($200-500/year potential!)

Next Module: Module 2: Prompt Engineering Fundamentals

In Module 2, you’ll learn:

The art and science of prompt engineering
How to structure prompts for best results
Few-shot learning techniques
Chain-of-thought prompting
Works with BOTH local AND API models!

Why Module 2 is critical: Master prompting → Get 3-5× better results from ANY model (local or API)!

Ready? Let’s master prompt engineering in Module 2!

Last updated: 2025-11-22 Module status: Complete Cost impact: $200-500/year savings potential Tools: Ollama, Aider, Continue.dev