Low-Cost Coding with Qwen 3.6

First ran Qwen3.6-35B-A3B (Q4) on a dual 3090 rig with 200k context, then switched to Qwen3.6-27B (Q8) the next day, and Claude Code tasks continued to work fine.

The setup followed the Unsloth quickstart to spin up llama-server, then connected the claude CLI via an Anthropic-compatible endpoint.

• --ctx-size 200000
• --temp 0.6, --top-p 0.95, --top-k 20, --min-p 0.00
• ANTHROPIC_BASE_URL=http://192.168.18.4:8001

Over 8 hours, Claude Code's cost estimate was $142, while local electricity cost was calculated at under $4. This calculation is a rough estimate assuming constant 1kW consumption, and the author noted actual power draw may have been lower.

The author put the rig build cost at about NZD 4,500, and based on this figure estimated the break-even point for using local instead of the Anthropic API at roughly 260 hours.

What was actually built was a Rust backend and an SSE web dashboard for monitoring server resources, with interaction limited to about 1 initial prompt plus 4 rounds of UI/UX and bug fixes.