Gemma 4 31B beats GPT-5.4-Pro

This is a systems win more than a raw-model victory: the post is really about orchestration, persistence, and memory compression turning a strong open model into something that can outlast a bigger baseline on a narrow task.

• –Google positions Gemma 4 31B for agentic workflows, coding, and tool use, so this result fits the model’s intended lane rather than contradicting it.
• –The repo’s three-agent loop is the real feature here: a main generator, an iterative critic, and a memory agent that condenses history when the context window gets tight.
• –The claim is still anecdotal until the task, prompts, success criteria, and token cost are published; “worked for 2 hours” can hide a lot of hand-tuning.
• –Community discussion around Gemma 4 has already surfaced memory-pressure and tool-call stability issues, so this reads like a proof of concept, not a blanket endorsement.
• –The broader takeaway is that open models may become much more competitive once you pair them with durable memory and correction loops instead of judging them as one-shot chatbots.