# Tuning Qwen 3.8 27B and OMP as a coding agent on 2× 3090s

Digest AI · Marketing & Small Business · published 2026-09-13T00:00:00Z

Canonical: https://digestai.news/story/tuning-qwen-3-8-27b-and-omp-as-a-coding-agent-on-2-3090s

## Summary

The hardware setup consists of two used RTX 3090 GPUs running vLLM. By default, omp is tuned for hosted cloud models, which can cause local setups to experience severe latency or truncated file writes.

To optimize performance, the developer adjusted five key settings in omp. These changes include setting explicit effort levels for all roles, establishing a thinking budget of 6,000 to 7,500 tokens, raising the maximum token limit to 32,768, saving tool outputs over 10 KB to files, and limiting parallel subagents to four.

## Key points

- Setting a thinking budget of 6,000-7,500 tokens prevented the model from wasting time on excessive reasoning.
- Raising maxTokens to 32,768 stopped the agent from saving half-written files due to truncated outputs.

## Why it matters

This guide shows how developers can run capable, private coding agents locally on consumer hardware, avoiding recurring cloud API fees and data privacy concerns.

## Sources

1. [Tuning Qwen 3.8 27B and OMP as a coding agent on 2× 3090s](https://doug.sh/posts/tuning-a-local-coding-agent-oh-my-pi) (doug.sh, 2026-09-13)

## Cite

Digest AI, "Tuning Qwen 3.8 27B and OMP as a coding agent on 2× 3090s", 13 September 2026, https://digestai.news/story/tuning-qwen-3-8-27b-and-omp-as-a-coding-agent-on-2-3090s

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/tuning-qwen-3-8-27b-and-omp-as-a-coding-agent-on-2-3090s.json