Plan With a Frontier Model, Code With a Local Quantized Qwen

Melvin Vivas · X post · 2026-08-30 · Open on X

Topics: AI Dev Tools & Productivity, LLM Fundamentals, LLMOps, Deployment & Monitoring · Level: intermediate

Summary

The creator thinks about a hybrid coding setup: use Codex with a frontier model (Sol) for planning, and hand the coding to Qwen3.8-27B quantized to Q4_K_M running locally. A local model has no usage limits, so it can run in a loop for free even if it is slower.

Key points

• Split the work: a strong hosted model does the planning, a local model does the coding.
• Qwen3.8-27B at Q4_K_M quantization is the local model being considered.
• Local inference has no rate or usage limits, so you can run agent loops without paying per token.
• The trade-off is slower coding in exchange for free, unlimited runs.
• The creator asked ChatGPT whether the local model compares to the hosted 'Luna' model, which is not a real benchmark.

Resources mentioned

• Qwen3.8-27B · tool · huggingface.co · free
  A 27B-parameter open-weight model from Alibaba's Qwen family that you can run locally or call through hosted APIs.
  Also in: Deploying Qwen3.8 27B on Hugging Face Inference Endpoints (Melvin Vivas on X · notes), Using Hugging Face credits: Jobs, Inference Endpoints and Open Models (Melvin Vivas on X · notes), Running Qwen3.8-27B Locally on an M5 Max MacBook with Inco Splash (Melvin Vivas on X · notes), Ternary Bonsai 2 27B: 9x smaller model keeping 98.2% of benchmark scores (Melvin Vivas on X · notes) and 24 more
• OpenAI Codex · tool · openai.com · paid
  OpenAI's coding agent. In the diagram it writes code, fixes review findings and drives the build loop. The creator also used it to make this video.
  Also in: An agent bot that installs and drives Codex on its own (Melvin Vivas on X · notes), Asking a Coder bot to install Codex (Melvin Vivas on X · notes), Sign in with ChatGPT: Setting Usage Limits for Each App (Melvin Vivas on X · notes), Codex Cloud Environments Must Be Saved & Published Before Use (Melvin Vivas on X · notes) and 240 more

Try this

• Build a planner/executor coding loop: a hosted frontier model writes the plan and a locally hosted quantized Qwen carries it out.