← sources

Best first model for high RAM, limited VRAM for coding

source: reddit (reddit:LocalLLaMA)published 2026-08-29 17:50ingested 2026-08-30 02:40

retained text

Hi all,

I'm looking to try out a local model for agentic coding (for now; seems like an easy starting place). I've used opencode and cloud-based open models for personal projects and I'm hoping to sell colleagues on local models. At home I only have a 2018 laptop so no possibility to try things out for personal use first. At work we have some reasonably beefy hardware that's underutilized, but it's mostly high RAM. Here are the specs for one workstation:

- GPU1: Nvidia GeForce 940MX, 12GB VRAM
- GPU2: Nvidia RTX A4500 20GB
- 256GB RAM

(Didn't note CPU, but I'm sure it's decent.)

Seems like llama.cpp is the way to go given the hardware. For model, I'm debating between getting my feet wet with a quantization of Qwen 3.8 27B that'll easily fit on the RTX A4500 VRAM vs something larger that can utilize the massive RAM but might be a push for my current skills. I'm a Linux guy but the computer is Windows and needs to stay that way so I'd default to doing things in WSL2.

This build seems relevant: [https://www.reddit.com/r/ollama/comments/1vzip5o/qwen38flashnext\_176b\_on\_a\_16gb\_card\_yes\_it\_works/](https://www.reddit.com/r/ollama/comments/1vzip5o/qwen38flashnext_176b_on_a_16gb_card_yes_it_works/)

Also, as a basic sanity check, am I going to get anything out of this hardware that's likely to produce useful outputs in an 8 hour workday with current models? Or am I just going to engender skepticism?

Thanks for any tips!

Edit: thanks all for the great suggestions!

document fetched 2026-08-30 00:37 · canonical https://reddit.com/r/LocalLLaMA/comments/1w1t3aw/best_first_model_for_high_ram_limited_vram_for