Qwen 3.8 27b - PI AGENT vs OPENCODE
https://www.reddit.com/r/LocalLLaMA/comments/1j7r47l/i_just_made_an_animation_of_a_ball_bouncing/
This post inspired me to make that test after a year ;)
That is one of my many tests I make comparing output quality.
What is more interesting using a PI Agent results are much better than an Opencode using a Qwen 3.8 27b ?!
Seems PI Agent is much better in the agent environment somehow... Not counting uses less tokens , do not have a hard limit of 32k output tokens, is faster, do not freezing, compressing context far less than Opencode. For instance if you have context in the Opencode output 32k and all context 100k then the compression is starting at 67k context ... PI is starting at 90k context even if you have set output context 64k or more.
My config for RTX 3090
llama-server with ini config -> which is exposing API to Opencode and PI agent.
llama-server.exe --models-preset 1_preset.ini --models-max 1 --direct-io
config ini
[Qwen3.8-27B_dense_c-100k]
model = models/Qwen3.8-27B-Q4_K_M.gguf
mmproj = models/mmproj-BF16-Qwen3.8-27B-UD-Q4_K_XL.gguf
reasoning-format = deepseek
flash-attn = on
n-gpu-layers = 99
reasoning = on
ctx-size = 100000
temperature=1.0
top-p=0.95
top-k=20
min-p=0.0
presence-penalty=0.0
repeat-penalty=1.0
mmproj-offload = false
ONE MORE IMPORTANT THING:
Always use a VISION module as the model is using vision to asses the output quality!
I am offloading it to a RAM as we do not need an extremely fast vision for a code.
A screenshot processing on a GPU 0.3s vs a RAM 3s do not make a big difference on a few screenshots during a code generation / debugging ;)
