Gemma passes 7 out of 8

After running Gemma 4 through 8 real work-use prompts, both 31B Dense and 26B A4B MoE showed fairly strong real-world performance.

In particular, 31B Dense passed 7 out of 8, including a test the author deliberately designed to fail, and was therefore judged to have come closer, for the first time, to being a free local LLM worth using for simple-to-medium-difficulty productivity tasks.

The tests weren't benchmarks but prompts constructed to be usable in actual work, and the following were released together with them.

• The original text of the 8 prompts
• The full model outputs for the long tests
• Source for a single HTML demo app that can be run with a free AI Studio key

Additional verification was also attached.

• Results were independently confirmed by Gemini 3.1 Pro and Claude Opus 4.6 respectively
• However, the author specified that this test was run using the Genai API of GCP-hosted Gemma 4, not locally
• An acquaintance commented that results from local 31B execution were similar

The key point is that this is a reproducible real-world evaluation, including the published prompts and outputs. It's a case that lets us gauge, at this point in time, not just the model's raw performance but whether open-weight models can actually be put into real work.