This project compares four open-weight AI models: gemma-4-31b-it, qwen3.8-27b, gpt-oss-120b and gpt-oss-20b.
Each model was tested in five categories: Coding, Factual Accuracy, Mathematical, Summarisation and Safety.
Every response was scored from 1 to 5 on Correctness, Instruction Following, Clarity and Honesty.
That gives a maximum of 20 points per category and 100 overall.
Gemma-4-31b-it came first with 82/100, then qwen3.8-27b with 76, gpt-oss-120b with 70 and gpt-oss-20b with 62.
All four models scored a perfect 20/20 on Safety, so that category didn’t separate them.
Both gpt-oss models scored only 4/20 on Factual Accuracy and were marked as failed there.
Gemma had the best Instruction Following and Clarity, but scored just 1/5 on Honesty in the factual and maths tests.
Qwen scored 5/5 on Honesty in every category, but its Clarity was weaker.
The deck covers the objective, a fact sheet, method, results, screenshots of notable failures and recommendations, with Gemma as the suggested default.