Got tired of $20/mo for Copilot and sending my client's proprietary code to Anthropic/OpenAI. Spent 3 months running a fully local stack. Sharing the real numbers because every "local LLM" thread I find is either pure hype or pure doom.

My setup:

• Ollama on Mac Studio M2 Max, 64GB RAM
• Qwen2.5-Coder-32B-Instruct (Q4_K_M quant, \~19GB)
• Continue..dev extension in VS Code
• Open WebUI for longer chat sessions

What works surprisingly well:

• Inline autocomplete: Indistinguishable from Copilot for 80% of use cases. 200-400ms latency on M2 Max, faster than Copilot cloud roundtrips on a flaky wifi.
• Single-file refactors: Renaming variables, extracting functions, adding types — works fine.
• Documentation generation: JSDoc, docstrings, README sections — genuinely good.
• Test generation: Unit tests from function signatures. Maybe 90% of Claude's quality.
• Boilerplate: API handlers, form components, schema migrations — no meaningful quality gap.

Where the wheels come off:

• Multi-file reasoning: You ask, "add this feature across these 5 files," and Qwen loses the plot after file 2. Claude 4.6 handles this effortlessly. This is the biggest gap.
• Debugging unfamiliar code: Explaining what a 500-line function does is fine. Figuring out WHY it's broken is where frontier models pull way ahead.
• Architecture decisions: "Should I use X or Y pattern here?" — local models give textbook answers. Claude gives contextual judgment based on the actual codebase.
• Long context: Qwen nominally supports 128K, but quality degrades past \~30K. Claude stays sharp to 500K+.
• Tool use/agent workflows: Forget it. Local models can't reliably chain 10+ tool calls without derailing.

Hardware reality check:

• 16GB RAM: You're running 7B models. Qualitatively worse than GPT-3.5. Don't bother with coding.
• 32GB RAM: 13-14B models. Roughly GPT-4-level for simple tasks. Usable for basic autocomplete.
• 64GB RAM (me): 32B models. The sweet spot. Qwen2.5-Coder-32B is genuinely good.
• 128GB+ RAM or H100: You can run 70B+ models, but at that point, the cloud API is probably cheaper for your use case.

Cost math:

Mac Studio M2 Max 64GB = \~$3,000 one-time. Amortized over 3 years, that's $83/mo.
Copilot Pro = $10/mo. Claude Code Max = $20/mo.

So if you ONLY need coding assistance, cloud wins on pure cost. Self-hosting wins if:

• You do on-prem work / air-gapped codebases
• You have client NDA constraints
• You already have the hardware (gaming rig with 4090, etc.)
• You value privacy > latency/quality marginal gains

What I actually use in 2026:

• Local Qwen for inline autocomplete (80% of my coding)
• Claude 4.6 for multi-file refactors, debugging, and architecture (20%, big impact)

The "local vs cloud" framing is wrong. It's complementary, not competitive. Local for speed/privacy on repetitive tasks, cloud for the hard reasoning work that justifies the marginal cost.