Gemma 4 MoE for Agentic Coding: Testing Open-Weight Models on AMD APU Hardware

2 min read

Alex Ewerlof has written up what it's actually like to use local models for coding work, prompted by cloud pricing moving past its introductory phase. He ran Gemma 4 variants on an AMD Ryzen 7 PRO 250 APU with 64GB of RAM, driving them through LM Studio with a 150k context window. The concrete result he reports is Gemma 4 26B MoE generating a working Snake game.

His conclusion is measured rather than triumphant: open-weight models underperform the flagship proprietary ones, but in his words, "with tooling (e.g. an AI harness) you can compensate a lot." The drawbacks he names are cold-start delay and slower inference than a hosted API — the familiar cost of keeping the model on your own machine.

Worth reading for the setup detail rather than the verdict. An APU with 64GB of shared memory is a different proposition from a discrete GPU, and the combination of that hardware, a 150k context and an agent harness is a specific configuration you can compare your own against. One generated game is a demonstration, not a benchmark, so treat the capability claims as a starting point for your own testing rather than a settled result.

Read the full article on blog.alexewerlof.com.


Source: blog.alexewerlof.com · Relevance: 8/10