530 Developers on HN Are Ditching Claude and GPT for Local Models. The Service Business: Help Teams Make the Switch Without Breaking Their Workflow.

A thread titled "Ask HN: Has anyone replaced Claude/GPT with a local model for daily coding?" just hit 530 points on Hacker News.

That is not a signal to dismiss.

That is a market telling you what it wants.

What the Thread Reveals

The developers in that thread are not hobbyists running experiments on the weekend.

They are people with real codebases, real deadlines, and real reasons to stop paying for cloud AI:

• Privacy: some teams cannot send code to an external API under any circumstances
• Cost: $200/month per developer adds up fast at team scale
• Lock-in: a Claude or GPT price change, capability regression, or API outage breaks their entire workflow

And they are making it work.

The community consensus has settled on Qwen 3.6 35B-A3B and Gemma 4 31B as the practical sweet spots. On a Mac Studio with 128GB RAM, a MacBook with 36GB, or a rig with dual RTX 3090s running around 150 tokens per second, these models are genuinely usable for daily coding.

One developer described the experience: "It's roughly where I felt Claude was a year back — most sessions need more pair programming than solo agent work."

That is not a condemnation. That is a product that is one year behind the frontier and already useful for scoped tasks.

The Problem Is Not the Model

The problem is the setup.

The thread also documents what gets in the way:

• Local models require more precise prompting than frontier models
• They "get into loops quite often" without the right scaffolding
• Complex architectural decisions still push teams back to cloud
• Context window management for large codebases is harder without a proper tooling layer

None of these are unsolvable. They are configuration and workflow problems. And configuration and workflow problems are service businesses.

The Three Things Teams Actually Need

There are three things a developer or team needs to switch from Claude to a local model and stay switched:

1. Hardware and model selection

Not every setup works. A team on MacBook Pros with 16GB RAM has a different answer than a team with a shared GPU machine. Getting this wrong means slow inference, poor output, and immediate abandonment. The right recommendation requires understanding their hardware, their codebase size, and their daily task mix.

2. Workflow integration

Local models do not arrive in VS Code or Cursor with a working plugin. They need an inference endpoint (Ollama is the standard), an IDE integration (Continue.dev), and a prompt setup tuned to the team's codebase structure. Without this layer, developers get the model running and then cannot make it useful.

3. Prompting and reliability tuning

The failure mode that kills local model adoption is the loop problem. A model that gets stuck in repetitive output or loses track of a task will get abandoned in the first week. Fixing this requires structured context injection, role framing, and step-by-step instruction design — work that takes hours to get right and most developers cannot spare those hours.

What to Sell

The Local AI Coding Stack Audit — 1-week fixed-scope engagement.

Deliverables:

• Hardware spec recommendation for their team's existing machines
• Model selection and benchmark comparison on 20 real tasks from their codebase
• Written setup guide: Ollama + Continue.dev + IDE integration, step by step
• Prompting cheatsheet for the specific failure modes their codebase is likely to trigger
• Side-by-side output comparison: local vs. cloud, so the team sees the gap before committing

Price: $1,500 to $3,000 flat depending on team size and codebase complexity.

The Onboarding Sprint — 2-week engagement added on top of the audit.

You work alongside the team on real coding tasks for two weeks: pairing, iterating on prompt templates, fixing the loop problem as it surfaces, and setting up fallback routing to cloud for the task types where local models still underperform.

Price: $3,000 to $6,000.

The Model Update Retainer — ongoing monthly.

Qwen ships updates. Gemma ships updates. Every new version changes the behavior the team's prompts were tuned for. Someone needs to evaluate the new version against the team's task suite, update the setup, and push the changes before the team notices a regression.

Price: $500 to $1,000/month.

Finding Clients

The thread itself is your lead list.

Any developer in that 264-comment thread who is describing setup problems, asking for hardware advice, or writing "I tried it but could not make it reliable" is a warm lead. They have the motivation. They are missing the setup expertise.

Your reach-out is not a pitch. It is a follow-up:

"I saw your comment in the HN thread about switching to Qwen 3 for coding. We run local stack audits that fix exactly the prompting and reliability issues you described. Happy to show you what that looks like for a codebase your size."

One sentence about what you saw. One sentence about what you do. One sentence offering to be concrete.

The developers who responded in that thread are already past the "is this possible" question. They are at the "how do I make it reliable" stage. That is where your service starts.

Source: https://news.ycombinator.com/item?id=48542100