Running a 27B local LLM on dual RTX 5060 Ti: real-world pair programming setup and prompting tips
Kahvana · reddit · 2026-09-08
The author shares a local setup running the 27B model (Unsloth Q4KS, 200K context Q40) for pair programming.
Setup
- Dual RTX 5060 Ti 16GB (PCIE 5.0 x8x8), llama.cpp with tensor parallel and MTP speculative decoding (draft max-n 11, draft min-p 0.1)
- Kilo Code (VSCode extension) as harness, default jinja template
- Runs xhigh reasoning by default, stepping down to medium/low for simple or speed-sensitive tasks
Experience
- Impressive as a pair programmer: catches small mistakes, usually picks the right options; weaker on UI designs that diverge from the norm
Prompting tips that worked
- Hand-crafted ASCII renditions of components/layouts beat example images
- Describing ideas via class names, fields/properties, and method signatures lands faster than prose
The author is collecting other users' prompting techniques for this model.
More from coding & agent
- OpenAI Veteran Counters Codex Origin Story: It Was Built as an Internal Infra Tool — gabrielchua · 2026-09-08
- GLM-5.3-Flash builds a 3D kitchen in Blender; author argues Blender-RL is not GPU-expensive — bookwormengr · 2026-09-08
- Factory's /missions pre-build planning wins users over: architecture, milestones, and a HTML-review trick — matanSF · 2026-09-08
- Rauchg launches OSS grants v2: $1,000 each to 35 contributors across agents, local AI, performance — iamsahaj_xyz · 2026-09-08
- MCP tooling explodes on PyPI: mcp-types up 425%, OpenHands doubles 4 months straight — rajistics · 2026-09-08
- Muse Spark 1.3 builds a 16 km² browser 3D driving demo in Three.js — alexandr_wang · 2026-09-08