im looking for an llm to act as the builder agent. a separate planner agent writes the strong plan using powerful API model, and the builder executes it as the plan (so to make coding in fast time and save moeny in the powerful api) hardware and software: CPU: i7 10700K GPU: GTX 1080 Ti 11GB System RAM: 32 GB ddr4 3600Mhz OS: Debian 13 Backend: llama.cpp with Vulkan. CUDA is unstable on my 1080 Ti. After a few tens of thousands of tokens it fails with a "no memory space" error or someting like this Requirements: Use case: coding and agentic coding Context size: like from 100k to 128k will be amazing Priorities: 1st quality, 2nd speed, 3rd size Models I've tried: Qwen3.6 35B A3B UD Q4_K_M slow speed like about 7 to 10 t/s. and I only ran it on a few small tasks, so I haven't tested it properly Qwen3.5-9B-UD-Q4_K_XL (MTP) was a good speep i think from 25 to 30 t/s but not usable in builder agent Sharp Spark X2.5 4B Q6_K_XL overthinks too much and unusable at all in builder agent situation Sharp Spark X2.6 4B Q6_K_XL: fine for simple everyday tasks but it's still not good enough for real work builder agent has anyone used a tiny or small or medium-large model at Q3 or Q4 as a builder agent in a real world coding tasks and gotten at least ok results? i'd like to hear which model, which quant, which harness, and what kind of tasks it handled. thanks   submitted by