Hello fellow members of this lovely community,

Let me start by saying that I’m about as far from a professional developer as it gets. I’m a hobbyist whose entire coding experience consists of building various Python/VBA tools and simple JavaScript web apps mostly using VS Code. So far, my approach to using AI for coding has basically been copying and pasting sections of my code into ChatGPT and asking for changes or additions as needed.

Since small local models seem to have improved quite a bit for coding, I decided to dip my toes into this whole “agentic coding” space I’ve been hearing about. Hardware-wise, I have a measly 2080 Ti with 22 GB of VRAM, in which I managed to fit Unsloth’s Qwen3.6-27B-UD-Q4_K_XL with 128k context at q8_0 KV using the parameters below, while getting around 20–22 tok/s.

"qwen3.6-27b-coder":
    cmd: |
      ${llama_server}
      --host 0.0.0.0 --port ${PORT} -ngl 999  -fa on  --jinja --no-mmap -cram 2048 --no-warmup -np 1 
      --model ${host_model_dir}/Qwen3.6-27B/Qwen3.6-27B-UD-Q4_K_XL.gguf
      --mmproj ${host_model_dir}/Qwen3.6-27B/mmproj-F16-Qwen3.6-27B.gguf
      --no-mmproj-offload
      --spec-type ngram-mod 
      --spec-ngram-size-n 24 
      --draft-min 12 
      --draft-max 48
      --ctx-size 131072
      --cache-type-k q8_0
      --cache-type-v q8_0
      --temp 0.6
      --presence-penalty 0.0
      --repeat-penalty 1.0
      --min-p 0.0
      --top-k 20
      --top-p 0.95
      --fit off
      --reasoning on
      --reasoning-budget -1
      --chat-template-kwargs '{"enable_thinking":true}'
      --chat-template-kwargs '{"preserve_thinking":true}'

While searching for a coding agent that fits my setup, I saw PI being recommended quite a bit for being fast and lightweight. I installed it, hooked it up with Qwen3.6, and so far so good.

The issue I’m running into is that PI feels like a very barebones “DIY” type of agent. I’m sure that’s great if you know what you’re doing, but as a complete beginner to CLI-based coding agents, I’m honestly a bit lost on how to use it effectively or what a good workflow even looks like.

So I have a few questions for you more knowledgeable folks:

• Should I stick with PI and just go through the documentation until I’m more comfortable? Or would it make more sense to switch to something more “batteries included” like Opencode, Qwencode, etc.? Alternatively, should I just stick with VS Code and use an extension that connects to a local LLM?

• Regarding my model choice: is 128k context and ~20 tok/s actually usable for coding, or would I be better off switching to a 35B MoE model with CPU offload for higher speed and/or context?

• Any recommended optimizations for my llama-server parameters?

• Lastly, I’m running into an issue with PI where, even though reasoning is enabled on the llama-server side, the model doesn’t seem to “think” based on my initial tests. The thinking_level setting in PI is also set to off, and I can’t seem to change it.

Thanks in advance for any help or guidance.