I Replaced Cloud AI with Local AI for 90 Days

Published on April 11, 2026 • 20 min read

On January 10th, I cancelled my ChatGPT Plus subscription, turned off GitHub Copilot auto-renewal, and stopped using Midjourney. For the next 90 days, every AI task would run on hardware I own, in my office, with no cloud API calls.

I wanted to answer three questions: Can local AI actually replace cloud AI for daily professional work? How much money does it save? And where does it fall apart?

Here is the full account. Not the polished version. The real one, including the days I almost gave up.

---

The Setup: Hardware and Software {#setup}

Hardware I started with:

• Desktop PC: Ryzen 7 5800X, 64GB DDR4, RTX 3090 24GB
• Total hardware cost when purchased: ~$1,200 (built in early 2025 with used parts)

Software stack:

• Ollama — Model management and inference engine
• Open WebUI — ChatGPT-like browser interface, connected to Ollama
• Continue.dev — VS Code extension for code completion and chat, connected to Ollama
• Flux (via ComfyUI) — Image generation, replacing Midjourney
• Whisper (large-v3) — Speech-to-text for meeting transcription

Models I loaded on day one:

ollama pull llama3.2:7b          # General purpose
ollama pull codellama:13b        # Code generation
ollama pull llama3.1:13b         # Longer context tasks
ollama pull deepseek-coder-v2:16b  # Code review

Cloud services I was replacing:

For the Open WebUI setup, I followed our own guide. Took about 15 minutes.

---

Week 1: The Honeymoon (Days 1-7) {#week-1}

Everything felt exciting. Open WebUI looks good. The chat interface is responsive. Llama 3.2 7B answers questions at 45 tok/s on the RTX 3090, which feels fast. I configured keyboard shortcuts in Continue.dev so that Tab-completion works like Copilot.

Day 1: Set up everything. Wrote a Modelfile for a "senior engineer" persona. Felt productive.

Day 3: Used local AI to draft four client emails, summarize a 12-page contract, and generate a Python ETL script. All tasks completed to my satisfaction. ChatGPT, who?

Day 5: First image generation with Flux. A promotional banner for a blog post. Took 45 seconds on the RTX 3090 (Midjourney does it in 15 seconds). Quality was good enough for web use but noticeably different from Midjourney's aesthetic.

Day 7 journal entry: "This is working better than expected. Speed is fine. Quality for daily tasks is sufficient. The privacy angle is real — I just pasted an entire client contract into the AI and got a summary without it leaving my machine."

Week 1 verdict: Optimistic. No major gaps yet.

---

Weeks 2-3: The Frustration Phase (Days 8-21) {#weeks-2-3}

Reality set in. The gaps between local and cloud AI are not about speed or interface. They are about intelligence.

Day 9 — The reasoning gap hits hard: I asked Llama 3.2 7B to analyze a complex database schema and suggest normalization improvements. ChatGPT-4o would nail this in one shot. Llama 3.2 7B gave me surface-level observations and missed obvious 3NF violations. I had to prompt it three times, breaking the problem into smaller pieces.

This became a pattern. Local models handle single-step tasks well. Multi-step reasoning — where the model needs to hold several constraints in mind simultaneously — is where 7B models visibly struggle compared to frontier cloud models.

Day 12 — Code completion frustration: Continue.dev with CodeLlama 13B is competent for boilerplate code. Autocomplete for common patterns works. But when I need it to understand the context of my project — the architecture, the naming conventions, the business logic — it falls short. Copilot had months of context from my repository. CodeLlama sees only what fits in its context window.

# My workaround: create a project context file and prepend it
cat > project-context.md << 'EOF'
Project: E-commerce API (Node.js/Express)
Database: PostgreSQL with Prisma ORM
Auth: JWT with refresh tokens
Naming: camelCase for JS, snake_case for DB columns
Error pattern: AppError class with status codes
EOF

# Then reference it in Continue.dev config

Day 15 — Image generation reality: Flux produces good images, but prompt engineering is different from Midjourney. I spent 40 minutes getting a simple product mockup that Midjourney would have generated from a one-line prompt. The aesthetic is also different. Flux tends toward photorealistic; I often wanted Midjourney's stylized look.

Day 18 — The knowledge cutoff wall: Asked the model about a library released two months ago. Blank stare. Cloud AI has web search plugins. My local model only knows what was in its training data. I started keeping a browser tab open alongside the AI chat, which partially defeats the purpose.

Day 21 journal entry: "Productivity has dropped about 20% for complex tasks. Simple tasks are identical speed. I'm spending more time crafting prompts and breaking problems into pieces. The model is not dumb, but it requires more guidance than ChatGPT-4o."

Weeks 2-3 verdict: Struggling. Considering quitting for complex work.

---

Month 2: Finding the Sweet Spot (Days 22-60) {#month-2}

Instead of trying to make local AI match cloud AI at everything, I got strategic about which tasks to assign locally.

The breakthrough: task-specific models.

I stopped using one model for everything. Each task got its own Modelfile with a custom system prompt, temperature, and sometimes a different base model.

# Coding assistant (low temperature, precise)
cat > Modelfile-coder << 'EOF'
FROM deepseek-coder-v2:16b
SYSTEM """You are a senior software engineer. Write clean, production-ready code.
Always include error handling. Prefer readability over cleverness.
If requirements are ambiguous, ask for clarification before coding."""
PARAMETER temperature 0.3
PARAMETER num_ctx 8192
PARAMETER top_p 0.85
EOF
ollama create coder -f Modelfile-coder

# Writing assistant (higher temperature, creative)
cat > Modelfile-writer << 'EOF'
FROM llama3.2:7b
SYSTEM """You are a professional writer. Write in a direct, clear style.
Avoid cliches, filler phrases, and corporate jargon.
Match the tone of the input: formal for business, casual for blog posts."""
PARAMETER temperature 0.8
PARAMETER num_ctx 4096
PARAMETER top_p 0.95
PARAMETER repeat_penalty 1.15
EOF
ollama create writer -f Modelfile-writer

# Data analyst (structured output)
cat > Modelfile-analyst << 'EOF'
FROM llama3.1:13b
SYSTEM """You analyze data and provide structured insights.
Always use tables, bullet points, or numbered lists.
Include specific numbers and percentages.
If data is insufficient for a conclusion, say so explicitly."""
PARAMETER temperature 0.4
PARAMETER num_ctx 8192
EOF
ollama create analyst -f Modelfile-analyst

Day 30 — The productivity recovery: Task-specific models plus better prompting habits recovered most of the lost productivity. I was not back to 100% compared to cloud AI, but I was at roughly 85-90% for my typical daily work.

Day 35 — Whisper transcription pays off: Had a 90-minute client meeting. Whisper large-v3 transcribed the entire recording in 12 minutes on my RTX 3090. Transcription quality was excellent — maybe 95% accurate including technical jargon. Then I fed the transcript to my analyst model for a summary with action items. The entire post-meeting workflow took 20 minutes instead of the usual hour of manual notes.

# Transcribe meeting recording
whisper meeting-recording.mp3 --model large-v3 --output_format txt --language en

# Feed to local AI for summary
cat meeting-transcript.txt | ollama run analyst "Summarize this meeting. List all decisions made, action items with owners, and open questions."

Day 42 — Continue.dev gets better with context: I set up a RAG pipeline using a local vector database to index my project's codebase. Continue.dev queries this index to find relevant code snippets and includes them in the context window. Code suggestions improved dramatically. Still not Copilot-level, but close enough for daily use.

For the Continue.dev + Ollama setup, the configuration took about 30 minutes including the RAG index.

Day 50 — Privacy as a feature, not a compromise: A colleague needed to analyze sensitive HR data for a compensation audit. On cloud AI, this would require legal review, DPA agreements, and probably a hard no from compliance. With local AI, we loaded the data, ran the analysis, and had results in 20 minutes. No data left the building. This is where local AI has an unbeatable advantage.

Month 2 verdict: Productive. Found the workflow. Not trying to replicate cloud AI, but using local AI where it excels.

---

Month 3: The Settled Workflow (Days 61-90) {#month-3}

By month three, I had stopped thinking about the experiment. Local AI was just how I worked. The tools were configured, the models were dialed in, and the friction was gone.

My daily routine became:

Day 68 — Model upgrade makes a difference: Pulled Qwen 2.5 32B (Q4_K_M quantization, fits in 24GB VRAM). The jump from 7B to 32B for general tasks was significant. Reasoning improved noticeably. Multi-step problems that stumped the 7B model were handled cleanly by the 32B. This model became my new default for anything beyond simple Q&A.

ollama pull qwen2.5:32b-instruct-q4_K_M
# 19GB download, loads fully into RTX 3090 24GB
# Generation speed: ~18 tok/s — slower than 7B but the quality jump is worth it

Day 75 — Flux workflow matures: After weeks of practice, my Flux prompting got efficient. I built ComfyUI workflows with saved presets for common image types: blog headers, social media posts, product mockups. Generation time per image dropped to 30 seconds with the right workflow. Quality was consistently good enough for professional use.

Day 82 — The one thing I still miss: Real-time web knowledge. When I need to research something current — a new API, a recently reported bug, a competitor's feature — local AI is useless. I keep a browser open and sometimes use Perplexity (free tier) for web-grounded answers. This is the one area where cloud AI has an insurmountable advantage over local models.

Day 90 journal entry: "Done. Not going back to full cloud. Not going full local either. Hybrid is the answer."

---

The Cost Breakdown: 90 Days of Numbers {#cost-breakdown}

Cloud AI costs I avoided:

Local AI costs incurred:

Net savings over 90 days: ~$102

The savings are real but modest. The bigger wins are not financial:

• Privacy: Processed sensitive client data, HR records, and financial docs without cloud exposure
• Availability: AI works during internet outages, on flights, in secure facilities
• No rate limits: Never hit a usage cap, never got throttled
• No subscription anxiety: The hardware is paid for; running costs are electricity only

Check our detailed cost comparison for different hardware tiers and usage patterns.

---

What Stuck: Tasks Where Local AI Won {#what-stuck}

1. Coding assistance (85% replacement) Continue.dev with DeepSeek Coder V2 16B handles 85% of what I used Copilot for. Tab completion, function generation, refactoring suggestions. The remaining 15% — complex cross-file refactoring and architecture suggestions — I handle manually now.

2. Document analysis and summarization (95% replacement) Summarizing contracts, reports, meeting transcripts. Local AI is perfect for this. The quality matches cloud AI for extractive tasks, and the privacy advantage is decisive.

3. Privacy-sensitive tasks (100% replacement) Anything involving client data, employee data, financial records, or proprietary code. Cloud AI was never an option for these tasks. Local AI made them possible.

4. Email and communication drafts (90% replacement) First drafts of emails, Slack messages, documentation. Llama 3.2 7B is fast and the quality is more than adequate for drafts that I edit before sending.

5. Meeting transcription (100% replacement) Whisper large-v3 is simply excellent. Transcription quality matches or exceeds cloud services, runs entirely offline, and processes a one-hour recording in about 8 minutes.

---

What I Went Back to Cloud For {#what-failed}

1. Complex multi-step reasoning Questions that require holding 5+ constraints simultaneously and reasoning through them step by step. "Design a database schema for X with these constraints, then generate the migrations, then write the seed data." GPT-4o and Claude still outperform even 32B local models here.

2. Latest knowledge Anything about events, releases, or changes from the last few months. Local models have training cutoffs. I use Perplexity free tier for web-grounded research.

3. Stylized image generation Flux is capable, but Midjourney has a specific aesthetic that clients sometimes request. For "make it look like a Midjourney image," I would need Midjourney. I now use image generation case-by-case: Flux for most things, Midjourney (re-subscribed at Basic $10/month) for client-facing work that demands that specific style.

4. Very long document processing Documents over 30,000 tokens push even 32B models to their limits. Context window constraints mean I have to chunk the document and process pieces separately, losing the ability to reference information across sections. Cloud models with 128K+ context windows handle this natively.

---

The Final Verdict: Hybrid Wins {#verdict}

After 90 days, here is where I landed:

Local AI handles 70-75% of my daily AI usage. Coding, summarization, drafts, transcription, privacy-sensitive work. These tasks run faster, cheaper, and more privately on local hardware.

Cloud AI handles 25-30%. Complex reasoning, current knowledge, very long documents, and specific image styles. I re-subscribed to ChatGPT Plus ($20/month) and Midjourney Basic ($10/month). Dropped Copilot entirely.

Monthly cost comparison:

Net monthly savings: ~$21/month, plus the privacy and availability benefits that have no price tag.

The money savings alone do not justify the switch. The real value is control. My AI tools work offline. They process sensitive data. They never change their pricing, deprecate features, or modify behavior with a silent model update. That stability is worth more than the $21/month.

For a deeper look at the numbers, see our local AI vs ChatGPT cost analysis and the free models guide for building your own stack.

---

If I Were Starting Over: What I Would Do Differently {#starting-over}

1. Start with the 32B model, not 7B. The quality jump is massive. If your hardware can run it, skip the 7B frustration phase.

2. Set up task-specific Modelfiles on day one. Do not use one model for everything. Spend 30 minutes creating purpose-built configurations.

3. Install Whisper immediately. Meeting transcription was the highest-value local AI use case from day one.

4. Do not try to fully replace cloud AI. Go hybrid from the start. Use local for 70% of tasks, cloud for the rest. The all-or-nothing approach wastes time.

5. Budget more time for Flux learning. Image generation has the steepest learning curve. Expect a week of experimentation before you are productive.

---

Conclusion

Ninety days taught me that the question is not "local AI or cloud AI?" It is "which tasks belong where?"

Local AI excels at private, repetitive, well-defined tasks. Cloud AI excels at novel, complex, knowledge-current tasks. Running both costs less than running cloud alone, and gives you capabilities that cloud alone cannot provide.

The experiment started as a challenge. It ended as a permanent workflow change.

---

Want to build your own local AI workflow? Our courses walk you through the entire setup, from hardware selection to production deployment, with hands-on labs for Ollama, Open WebUI, Continue.dev, and more.